CreatingValuewithBigDataAnalytics

Ournewlydigitalworldisgeneratinganalmostunimaginableamountofdataaboutallofus.Suchavastamountofdataisuselesswithoutplansandstrategiesthataredesignedtocope with its size and complexity, and which enable organ-isations to leverage theinformation to create value. This book is a refreshingly practical yet theoretically soundroadmaptoleveragingbigdataandanalytics.

Creating Value with Big Data Analytics provides a nuanced view of big datadevelopment, arguing that big data in itself is not a revolution but an evolution of theincreasing availability of data that has been observed in recent times. Building on theauthors’extensiveacademicandpracticalknowledge,thisbookaimstoprovidemanagersandanalystswith strategicdirectionsandpracticalanalytical solutionsonhowtocreatevaluefromexistingandnewbigdata.

Bytyingdataandanalyticstospecificgoalsandprocessesforimplementation,thisisamuch-needed book that will be essential reading for students and specialists of dataanalytics,marketingresearch,andcustomerrelationshipmanagement.

PeterC.Verhoef is Professor ofMarketing at theDepartment ofMarketing, Faculty ofEconomics and Business, University of Groningen, The Netherlands. He also holds avisitingprofessorshipinMarketingatBINorwegianBusinessSchoolinOslo.

EdwinKooge is co-founder ofMetrixlab Big Data Analytics, The Netherlands. He is apragmaticdata analyst, a result-focused consultant, andentrepreneurwithmore than25years’experienceinanalytics.

NatashaWalk is co-founderofMetrixlabBigDataAnalytics,TheNetherlands. She is adata hacker, analyst, and talent coach with more than 20 years’ experience in appliedanalytics.

Thisisatimelyandthought-provokingbookthatshouldbeonamust-readlistofanyoneinterestedinbigdata.

SunilGupta,EdwardW.CarterProfessorofBusiness,HarvardBusinessSchool,USA

This isoneof themostcompellingpublicationson thechallengesandopportunitiesofdataanalytics. Itpaintsnotonlyatheoreticalframework,butalsonavigatesmarketingprofessionalsonorganizationalchangeanddevelopmentofskillsandcapabilitiesforsuccess.Amust-readtounlockthefullpotentialofdata-drivenandfact-basedmarketing!

HarryDekker,MediaDirector,UnileverBenelux,TheNetherlands

CreatingValuewithBigDataAnalyticsoffersauniquelycomprehensiveandwell-groundedexaminationofoneofthemost critically important topics in marketing today.With a strong customer focus, it provides rich, practicalguidelines,frameworksandinsightsonhowbigdatacantrulycreatevalueforafirm.

KevinLaneKeller,TuckSchoolofBusiness,DartmouthCollege,USA

Nolongercanmarketingdecisionsbemadeonintuitionalone.Thisbookrepresentsanexcellentformulacombiningleadingedgeinsightandexperienceinmarketingwithdigitalanalyticsmethodsandtoolstosupportbetter,fasterandmore fact-based decision-making. It is highly recommended for business leaders who want to ensure they meetcustomerdemandswithprecisioninthe21stcentury.

MortenThorkildsen,CEORejlers,Norway;chairmanofITandcommunicationscompany,Itera;formerCEO,IBMNorway(2003–13);ex-

chairmantheNorwegianComputerSociety(2009–13),andvisitinglecturerNorwegianBusinessSchool,Norway

BigDataisthenextfrontierinmarketing.Thiscomprehensive,yeteminentlyreadablebookbyVerhoef,KoogeandWalk isan invaluableguideandamust-readforanymarketerseriously interested inusingbigdata tocreate firmvalue.

Jan-BenedictE.M.Steenkamp,MasseyDistinguishedProfessorofMarketing,MarketingAreaChair&ExecutiveDirectorAiMark,Kenan-Flagler

BusinessSchool,UniversityofNorthCarolinaatChapelHill,USA

Thisbookgoesbeyondthehype,toprovideamorethoroughandrealisticanalysisofhowbigdatacanbedeployedsuccessfullyincompanies;successfulinthesenseofcreatingvaluebothforthecustomeraswellasthecompany,aswellaswhatthepre-requisitesaretodoso.Thisbookisnotaboutthehype,norabouttheanalytics,itisaboutwhatreallymatters:howtocreatevalue.Itisalsoillustratedwithabroadrangeofinspiringcompanycases.

HansZijlstra,CustomerInsightDirector,AIRFRANCEKLM,TheNetherlands

CreatingValuewithBigDataAnalytics

MakingsmartermarketingdecisionsPeterC.Verhoef,EdwinKoogeandNatashaWalk

Firstpublished2016

byRoutledge

2ParkSquare,MiltonPark,Abingdon,OxonOX144RN

andbyRoutledge

711ThirdAvenue,NewYork,NY10017

RoutledgeisanimprintoftheTaylor&FrancisGroup,aninformabusiness

©2015PeterC.Verhoef,EdwinKoogeandNatashaWalk

TherightofPeterC.Verhoef,EdwinKoogeandNatashaWalktobeidentifiedasauthorsofthisworkhasbeenasserted

bytheminaccordancewithsections77and78oftheCopyright,DesignsandPatentsAct1988.

Allrightsreserved.Nopartofthisbookmaybereprintedorreproducedorutilisedinanyformorbyanyelectronic,

mechanical,orothermeans,nowknownorhereafterinvented,includingphotocopyingandrecording,orinany

informationstorageorretrievalsystem,withoutpermissioninwritingfromthepublishers.

Everyefforthasbeenmadetocontactcopyrightholdersfortheirpermissiontoreprintmaterialinthisbook.The

publisherswouldbegratefultohearfromanycopyrightholderwhoisnothereacknowledgedandwillundertaketo

rectifyanyerrorsoromissionsinfutureeditionsofthisbook.

Trademarknotice:Productorcorporatenamesmaybetrademarksorregisteredtrademarks,andareusedonlyfor

identificationandexplanationwithoutintenttoinfringe.

BritishLibraryCataloguinginPublicationData

AcataloguerecordforthisbookisavailablefromtheBritishLibrary

LibraryofCongressCataloginginPublicationData

Verhoef,PeterC.,author.

Creatingvaluewithbigdataanalytics:makingsmartermarketingdecisions/

PeterVerhoef,EdwinKoogeandNatashaWalk.

pagescm

Includesbibliographicalreferencesandindex.

1.Consumerprofiling.2.Bigdata.3.Marketing–Dataprocessing.I.Kooge,

Edwin.II.Walk,Natasha.III.Title.

HF5415.32.V4752016

658.8’3–dc23

2015027898

ISBN:978-1-138-83795-9(hbk)

ISBN:978-1-138-83797-3(pbk)

ISBN:978-1-315-73475-0(ebk)

TypesetinBembo

bySunriseSettingLtd,Paignton,UK

To:Petra,AnneMiekeandMaurice

Contents

Listoffigures

Listoftables

Foreword

Preface

Acknowledgements

Listofabbreviations

1Bigdatachallenges

Introduction

Explosionofdata

Bigdatabecomethenorm,but…

Ourobjectives

Ourapproach

Readingguide

2Creatingvalueusingbigdataanalytics

Introduction

Bigdatavaluecreationmodel

Theroleofculture

Bigdataanalytics

Frombigdataanalyticstovaluecreation

Valuecreationmodelasguidanceforbook

Conclusions

2.1Value-to-customermetrics

Introduction

Marketmetrics

Newbigdatamarketmetrics

Brandmetrics

Newbigdatabrandmetrics

Customermetrics

Newbigdatacustomermetrics

V2Smetrics

ShouldfirmscollectallV2Cmetrics?

Conclusions

2.2Value-to-firmmetrics

Introduction

Marketmetrics

Brandmetrics

Customermetrics

Customerlifetimevalue

Newbigdatametrics

MarketingROI

Conclusions

3Data,dataeverywhere

Introduction

Datasourcesanddatatypes

Usingthedifferentdatasourcesintheeraofbigdata

Datawarehouse

Databasestructures

Dataquality

Missingvaluesanddatafusion

Conclusions

3.1Dataintegration

Introduction

Integratingdatasources

Dealingwithdifferentdatatypes

Dataintegrationintheeraofbigdata

Conclusions

3.2Customerprivacyanddatasecurity

Introduction

Whyisprivacyabigissue?

Whatisprivacy?

Customersandprivacy

Governmentsandprivacylegislation

Privacyandethics

Privacypolicies

Privacyandinternaldataanalytics

Datasecurity

Conclusions

4Howbigdataarechanginganalytics

Introduction

Thepowerofanalytics

Differentsophisticationlevels

Generaltypesofmarketinganalysis

Strategiesforanalyzingbigdata

Howbigdatachangesanalytics

Genericbigdatachangesinanalytics

Conclusions

4.1Classicdataanalytics

Introduction

Overviewofanalytics

Classic1:Reporting

Classic2:Profiling

Classic3:Migrationanalysis

Classic4:Customersegmentation

Classic5:Trendanalysismarketandsalesforecasting

Classic6:Attributeimportanceanalysis

Classic7:Individualpredictionmodels

Conclusions

4.2Bigdataanalytics

Introduction

Bigdataarea1:Webanalytics

Bigdataarea2:Customerjourneyanalysis

Bigdataarea3:Attributionmodeling

Bigdataarea4:Dynamictargeting

Bigdataarea5:Integratedbigdatamodels

Bigdataarea6:Sociallistening

Bigdataarea7:Socialnetworkanalysis

Emergingtechniques

Conclusions

4.3Creatingimpactwithstorytellingandvisualization

Introduction

Failurefactorsforcreatingimpact

Storytelling

Visualization

Choosingthecharttype

Conclusions

5Buildingsuccessfulbigdatacapabilities

Introduction

Transformationtocreatesuccessfulanalyticalcompetence

BuildingBlock1:Process

BuildingBlock2:People

BuildingBlock3:Systems

BuildingBlock4:Organization

Conclusions

6Everybusinesshas(big)data;let’susethem

Introduction

Case1:CLVcalculationforenergycompany

Case2:Holisticmarketingapproachbybigdataintegrationataninsurancecompany

Case3:Implementationofbigdataanalyticsforrelevantpersonalizationatanonlineretailer

Case4:Attributionmodelingatanonlineretailer

Case5:Initialsocialnetworkanalyticsatatelecomprovider

Conclusions

7Concludingthoughtsandkeylearningpoints

Concludingthoughts

Keylearningpoints

Index

Figures

1.1EffectsofnewdevelopmentsincludingbigdataonGDP

1.2Readingguideforbook

2.1Bigdatavaluecreationmodel

2.2Value-to-customervs.value-to-firm

2.3ClassificationofV2CandV2Fmetrics

2.4Bigdatavaluecreationmodellinkedtochapters

2.1.1Searchresultson“tablet”worldwide

2.1.2Searchinterestin“bigdata”and“marketresearch”

2.1.3Exampleoftrackingaidedandspontaneousawarenessthroughtime

2.1.4Exampleofbrandpreferenceofsmartphoneusers,de-averagedtogenderandage

2.1.5Brand-AssetValuator®model

2.1.6AssociationnetworkofMcDonald’sbasedononlinedata

2.1.7Averagenumberoflikesandcommentsperproductcategory

2.1.8Developmentofintimacyandcommitmentovertime

2.2.1UKsmartphonesales

2.2.2Exampleofbrandswitchingmatrix

2.2.3Brandrevenuepremium

2.2.4Relationshiplifecycleconcept

2.2.5TheCLVmodel:theelementsofcustomerlifetimevalue

2.2.6ExampleofgrossCLVdistributionperdecile

2.2.7CustomerequityROImodel

2.2.8Customerengagementvalue:ExtendingCLV

2.2.9ExampleofROIcalculation

3.1Twodimensionsofdata:Datasourceversusdatatype

3.2ExampleofNielsen-ClaritasinformationforaNewYorkZIP-code

3.3Illustrationofstructuredandunstructureddata

3.4ExampleofmarketdataonthesupplysideforUKsupermarkets

3.5Exampleofmarketdataonthedemandside

3.6Illustrationofbrandsupplydataextractedfrominternalsystems

3.7Illustrationofbranddemandbasedonmarketresearch

3.8Illustrationofadatamodelofcustomersupplydata

3.9Illustrationofcustomerdemanddata(NPS)

3.10The5“W”smodelforassessmentofdatasources

3.11Exampleofsimpledatatablewithcustomerascentralelement

3.12Exampleofproductdatatablederivedfromcustomerdatabase

3.13Netbenefitsofinvestingindataquality

3.1.1TheETLprocess

3.1.2Thedifferentdatatypes

3.1.3OverviewofsegmentationschemeusedbyExperianUK

3.1.4ExternalprofilingusingZIP-codesegmentationforclothingretailer

3.1.5PresenceofdatatypesforDutchfirms

3.1.6Thechallengesofdataintegration

3.2.1Dataprotectionlawsaroundtheglobe

3.2.2EffectivenessincreaseofFacebookadvertisingcampaignsafteradditionofprivacybutton

3.2.3Differentwaysofhandlingprivacysensitivedata

4.1Associationsbetweencustomeranalyticsdeploymentandperformanceperindustry

4.2Differentlevelsofstatisticalsophistication

4.3Optimizationofmarketsharevs.revenueperpricelevel

4.4Classificationofanalysistypes

4.5Bigdataanalysisstrategies

4.6Problem-solvingprocess

4.7Churnmodelresultsfortelecomfirm

4.8Tesco’sbeeranddiaperdata

4.9Differentconversionratesafterdeviceswitching

4.10Howbigdataarechanginganalytics

4.11ImpactofWhatsAppusageonthesmartphoneusageofaDutchtelecomcompany

4.12Caseexampleofmulti-sourcedataanalysisofrelationbetweenbrandperformanceandsalesshare

4.13Differenttypesofdataapproaches

4.14Averagetop-decileliftsofmodelestimatedattime

4.1.1Differentdistributionscausingsimilaraverages

4.1.2Exampleoftimeseriesforsales

4.1.3Profilingnewcustomersonageclassification

4.1.4Decileanalysisformonetaryvalueandretentionrates

4.1.5Gainchartanalysisforbookclub

4.1.6ExternalprofilingforaclothingretailerusingZipcodesegmentation

4.1.7Salessharepercustomersegmentfortotalcoffeeandfairtradecoffee

4.1.8Fallingsubscriptionbaseforatelecomprovider

4.1.9Decomposingsubscriptionbaseinacquisitionandchurn

4.1.10Migrationmatrixofcustomersofatelecomfirm

4.1.11Like-4-likeanalysisforvaluedevelopmentofthecustomerbaseofaphoneoperator

4.1.12StepsforexecutionofanL4Lanalysis

4.1.13Exampleofacohortanalysis

4.1.14Exampleofasurvivalanalysis

4.1.15Exampleofadendrogram

4.1.16Visualizationclusters

4.1.17Exampleofaclusteranalysisofshoppers

4.1.18Trendanalysis

4.1.19Effectsofdifferentmarketinginstrumentsonsalesforachocolatebrand

4.1.20PredictionsforservicequalitytimeseriesofaEuropeanpublictransportfirm

4.1.21Effectsofstoreattributesonstoresatisfaction

4.1.22Attributeschosenforstudyoncabservices

4.1.23Exampleofachoice-basedconjointdesignforacabstudy

4.1.24Segmentationanalysisforconjointstudyoncabservices

4.1.25ResponseratefordifferentRFM-segments

4.1.26ExampleofadecisiontreeusingCHAID

4.1.27OutputoflogisticregressionmailingexampleinSPSS

4.1.28Gainschart

4.2.1Onlinepurchasefunnel

4.2.2A/Btesting

4.2.3Effectofdifferenttouchpointsonadvertisingrecallandbrandconsideration

4.2.4Useofdifferentchannelforsearchandpurchase:Webroomingvs.showrooming

4.2.5Latentclasssegmentationbasedoncustomerchannelusage

4.2.6Revenues,costs,andprofitpergroupwithandwithoutsearchchannelcatalog

4.2.7Purchasefunnel:Pathtopurchaseonmobilehandset

4.2.8Comparisonofeffectsestimatedbyattributionmodelandlastclickmethod

4.2.9Closed-loopmarketingprocess

4.2.10Schematicoverviewofrecommendationagentinhotelindustry

4.2.11FluactivityUSApredictedbyGoogle

4.2.12Estimationresultsofmulti-levelmodeltoassessperformanceofCFMs

4.2.13Effectsofmarketingmixvariablesonbrandperformanceusingtime-varyingparametermodels

4.2.14Textanalyticsapproach

4.2.15IllustrationofPOStagging

4.2.16Illustrationofawordcloud

4.2.17Numberoftweetsbytimeandsentiment

4.2.18Degreecentrality

4.2.19Betweennesscentralityandclosenesscentrality

4.3.1Informationoverload

4.3.2Sweetspotofdata,storyandvisual

4.3.3Buildingblocksforaclearstoryline

4.3.4Analysisprocessvs.effectivecommunication

4.3.5Examplesofdifferentstorylinesfordifferentpurposes

4.3.6GraphofAnscombe’sQuartetdatatable

4.3.7Thepicturesuperiorityeffect

4.3.8Relationshipcharts

4.3.9Comparisoncharts

4.3.10Exampleofabulletchart

4.3.11Compositioncharts

4.3.12Distributioncharts

4.3.13Chartsuggestions—athoughtstarter

4.3.14Pre-attentiveattributes

4.3.15Basicanalyticalpatterns

4.3.16Fromstorylinetovisualstopresentation

5.1Shortageofsupplyinanalyticaltalent

5.2Changingroleofthemarketingintelligencedepartment

5.3Phasesofthestandardanalyticalprocess

5.4Multi-disciplinaryskillsofananalyst

5.5Possiblebigdatastaffprofiles

5.6Stepwisedevelopmentofanalyticalcompetencewithinthefirm

5.7Numberofvendorsinmarketingtechnologylandscaperepresentedinsupergraphicsofchiefmaric.com

5.8Differentlayersofabigdataanalyticalsystem

5.9Linkingdata,analyses,actionsandcampaigns

5.10FlowdiagramoftheadaptivepersonalizationsystemdevelopedbyChung,RustandWedel(2009)

5.11Organizationmodelsfortheanalyticalfunction

5.12Differentpersonalityprofilesofanalystsandmarketeers

6.1Valuedriversforanenergycompany

6.2ContributionofeachofthevaluedriverstoCLV

6.3ImpactofdifferentvaluedriverimprovementsonCLV

6.4Thebigdatadashboard

6.5Theconceptualmodelfortheholisticapproach

6.6Fromsearch/purchasebehaviortoproductcombinations

6.7Algorithmforcalculatingproductrecommendationsbasedontheproductrelationscore

6.8MapReduceprogrammingmodel

6.9Resultsofnewwayofworking

6.10Visualizationofmodelbeingused

6.11Comparisonofeffectsforattributionmodelandlast-clickmethod

6.12Comparisonofcomplexmodelwithsimplermodel

6.13Resultsofclusteranalysisonsocialnetworkvariablesoftelecombrand

7.1Keylearningpointsbychapter

7.2Wordcloudofourbook

Tables

2.1.1ExampleofitemsusedtomeasureRogers’adoptiondrivers

2.1.2DefinitionsofBAV®components

2.1.3Overviewofdifferentcustomerfeedbackmetrics

2.1.4Conceptualizationofcustomerfeedbackmetrics

2.1.5Criteriaforgoodmetrics

4.1.1Sevenclassicdataanalytics

4.1.2Gainsandliftscale

4.2.1Sevenbigdataanalytics

4.2.2HowInternetchoicediffersfromsupermarketchoice

5.1Shiftingfocusofthemarketingintelligencefunction

Foreword

Companiesaround theworldare strugglingwithavastamountofdata,andcan’tmakesense of it all. “Big data” has the promise of providing firms with significant newinformation about their markets, their products, their brands, and their customers—butcurrently,there’softenagreatdividebetweenbigdataandtrulyusableinsightsthatcreatevalueforthefirmandthecustomer.

Thisbookaddressesthishugeneed.WhenIhadtheopportunitytoreadCreatingValuewithBigDataAnalytics:Makingsmartmarketingdecisions,myfirstreactionwas:Thankgoodness!Wherehasthisbookbeenallmylife?Finally,here’sabookthatprovidesaclear,detailed,andusableroadmapforbigdataanalytics.Iknowthat’shardtobelieve,butreadon.

AsIwritethis,Facebookhasreachedanewmilestoneof1billionusersinasingleday.Just thinkof thebigdata analyticsopportunities from just thatoneday.Verhoef,Koogeand Walk have developed a theoretically sound and highly practical framework. Theirvalue creationmodel just makes sense; it makes the complex simple. First, they clearlyidentify the goal of any analytic “job to be done”, focusing on either (a) creating andmeasuringvalue to the customer, or (b) creating andmeasuring value to the firm. Theyfurther break these two goals down into three levels: market level, brand level andcustomer level. This clear delineation of six key analytic areas of focus, followed bypractical,“how-to”guidesforusingandanalyzingbigdatatoanswerquestionsineachofthese key areas, is a highly executable approach, well grounded in rigorous scientificresearch.

Theydoagreatjobofachievingthreekeyobjectives:

1. Teachingusallhow“bigdata”providenewopportunitiestocreatevalueforthecustomer(socustomerslikeourproductsandservicesbetter),andforthefirm(sowemakemore profit),while also helping us to bemindful of key security andprivacyissues.Thisframeworkmakesthebookwork.

2. Teaching us specific analytic approaches that truly fit identifiable marketingquestionsandsituations,and,mostimportantly,howtogaininsightsthatleadtovalue creation opportunities—new growth opportunities, new customers, orgrowthfromexistingcustomers.This isthemissingpiecethatthisbookdoessowell. One key advantage of this book is that it offers in-depth key analytic

approaches for all areas ofmarketing, including analytic classics, new big datatechniques,story-tellingandvisualization.

3. Teaching us how to develop a big data analytics capability focused on valuecreation—thatdeliversgrowthandpositiveROI.Bytakingus throughtheentireprocess from getting the data, to integrating the data, to analysis, to insight, tovalue, to the role of the organization—the roadmap is complete, and ready foranyonetobegin.

Whoshouldreadthisbook?Anyonewhoneedstounderstandcustomers,products,brands,marketsorfirms.CMOsandmarketingexecutivesshouldreadthisbook—itprovidesgreatinsights intohowyoucandevelopa successfulbigdataanalytics capability, andhow tointerpret insights from big data to fuel growth. Those individuals chargedwith insightswithin the organization should read this book: one of the key learnings from Verhoef,Kooge and Walk’s approach is that you’ll know what analysis to do, when, for whatpurpose, and with what data. That’s huge! Data scientists should read this book—notbecauseyouneed to learn the analysis techniquesdescribedhere (youmaybe awareofmanyof them),butbecause itwill strengthenyourability togain insightsonmarketingproblems and help you to communicate your ideas and insights to the rest of theorganization.Evenprofessorsandstudentsofanalyticsshouldreadthisbook.Itprovidesarigorous approach to frame your thinking and build your analytic skills. And finally, ifyourheadisswimmingandyou’reoverwhelmedwiththeopportunitiesandcomplexitiesofthe“firehose”ofbigdata,thisbookisforyou.

I believe it’s the Rosetta Stone we’ve all been looking for, finally answering criticalquestions:Howdowecreateinsightsfrombigdataformarketing?Howdowecreatevaluefrombigdata?Howdowesolveproblemswithbigdata?AndhowdowegetapositiveROI on our investment in big data analytics? Whether you are just starting on yourjourneyinbigdataanalytics,orwellonyourway,youwilllearnatonfromthisbook.

Theauthorsdon’tshyawayfromallthecomplexitiesandthemessinessofbigdataandanalytics. Rather, theymake the complexmanageable and understandable. They explaindifficult analytic approaches clearly and show you when— and why—to use whattechnique. They provide a rare combination of science and practicality. Examples, casesand practical guidelines are clear, detailed and readable, taking you to that next step ofgettingtothebusinessofanalyzingyourownbigdatatocreatevalueforyourcustomersandyourfirm.

WhatmorecanIsay?CreatingValuefromBigDataAnalytics:Makingsmartmarketingdecisionsoffersin-depth,rigorousandpracticalknowledgeonhowtoexecuteasuccessfulbigdataanalyticsstrategythatactuallycreatesvalue.Thisisthefirstbookthatputsitalltogether.Thanks somuch toPeter, Edwin andNatasha forwriting the book thatwe allreallyneeded.

KatherineN.Lemon,PhD

AccentureProfessorandProfessorofMarketing,CarrollSchoolofManagement,BostonCollegeExecutiveDirector,

MarketingScienceInstitute(2015–2017)

Preface

When we started our careers in marketing analytics, it was a small discipline whichattracted only minor attention from the boards of companies. Analytics was mainlydevelopedinfirmshavingastrongdirectmarketingfocus,suchasReadersDigest.Beyondthat,researchagenciesweretryingtodevelopanalyticalsolutionsformorebrand-orientedcompanies. During our careers this situation has dramatically changed. Analytics havebecome a major discipline in many firms and scientific evidence strongly supports theperformanceimpactofastronganalyticsdepartment.Successfulexamplesinleadingfirmsprovideonlymoresupportforhavingastronganalyticalfunction.Marketinghasbecomemoredata-driveninthepastdecade!

Thisdevelopmenthasonlybecomemoreprominentwiththearrivalof“bigdata”.CEOsofbanks, retailers, telecomproviders,etc.nowconsiderbigdataasan importantgrowthopportunityinseveralaspectsoftheirbusinesses.Despitethis,weobservethatmanyfirmsfacestrongchallengeswhendevelopingbigdatainitiatives.Manyfirmsembracebigdatawithout having a decent developed analytical function and without having sufficientknowledge in the organization on data analytics, let alone on big data analytics. Wethereforebelievetherewasanurgentneedtowriteabookoncreatingvaluewithbigdataanalytics.Insodoing,westronglysympathizedwiththeviewthattheexistenceofbigdatashouldnotbeconsideredarevolution;itratherbuildsonthestrongdevelopmentsindataandanalyticsinthepast.

It was not just external big data developments that led us to write this book: someinternalmotivations inducedusaswell.Allofus,atsomepoint inourcareerswhenwehad built up extensive knowledge on marketing analytics, felt the need to share thisknowledge with a broader audience, rather than only clients, fellow academics, and/orstudents.Wehadalreadydevelopedmaterialformasterstudentsandexecutivesinspecificspecializedprograms,suchasmasterclassesoncustomervaluemanagementandexecutiveprograms on customer centric strategies. However, whenwriting this book, we realizedthat thisknowledgewasnotsufficient.Theworldofbigdatahascreatednewanalyticalapproachesthatwehadtodiveinto.Moreover,thesedevelopmentsinspiredustorethinkourconceptsanddevelopnewframeworks.Overall,writingthisbookwasagreatlearningexperience forall ofus.Wehope thatyouwillhavea similar learningexperiencewhenyoureadthisbook.

Acknowledgements

Writing this book would not have been possible without the support of many people.ForemostwewanttothankKimLijdingwhogaveusconsiderablehelpinthefinalstagesof the book, especially in getting the chapters organized.We also thankHans RisseladaPhD for some collegial feedback and the many marketing managers and marketingintelligence managers who provided valuable input for our book in the developmentprocess.Wealso acknowledge the support fromNicolaCupit fromRoutledgeduring thewriting process. Finally, we want to emphasize that writing this book was a great andstimulatingjointexperience.Soenjoy!

PeterC.Verhoef,EdwinKoogeandNatashaWalk

Abbreviations

ANCOVA AnalysisofcovarianceANOVA AnalysisofvarianceAPE AveragepredictionerrorAPS AdaptivepersonalizationsystemsARMAX AutoregressivemovingaveragewithxvariablesARPU AveragerevenueperuserATL AbovethelineB2B Business-to-businessB2C Business-to-customerBAV BrandassetvaluatorBE BrandequityBRIC BrazilRussiaIndiaChinaBTL BelowthelineC2C Customer-to-customerCBC ChoicebasedconjointCDR CalldetailrecordCES CustomereffortscoreCFM CustomerfeedbackmetricsCHAID Chi-squareautomaticinteractiondetectionCIV CustomerinfluencevalueCKV CustomerknowledgevalueCLM ClosedloopmarketingCLV CustomerlifetimevalueCMO ChiefmarketingofficerCOGS CostsofgoodssoldCPC CostperclickCPO CostperorderCRM CustomerrelationshipmanagementCRV CustomerreferralvalueCSR CorporatesocialresponsibilityCTR ClickthroughrateCVM Customervaluemanagement

DASVAR DoubleasymmetricvectorautoregressiveDSI DigitalsentimentindexEBITDA Earningsbeforeinterest,taxes,depreciationandamortizationETL ExtracttransformloadeWOM Electronicword-of-mouthFMCG Fast-movingconsumergoodsFTC FederalTradeCommissionGDP GrossdomesticproductGMOK GeneralizedmixtureofKalmanfiltersmodelGRPs GrossratingpointsIT InformationtechnologyKPI KeyperformanceindicatorLP LoyaltyprogramMAPE MeanabsolutepercentageerrorMI MarketingintelligenceMSE MeansquarederrorNBD NegativebinomialdistributionNLP NaturallanguageprocessingNPS NetpromoterscoreNSA NationalSecurityAgencyOLAP OnlineanalyticalprocessingPCA PrincipalcomponentanalysisPOS Point-of-salePOST Part-of-speech-taggingPSQ PerceivedservicequalityRE RelationshipequityRFM RecencyfrequencymonetaryvalueROI ReturnoninvestmentSBU StrategicbusinessunitSEO SearchengineoptimizationSKU Stock-keepingunitTAM TechnologyacceptancemodelUGC UsergeneratedcontentUSP UniquesellingpointV2S Value-to-societyVAR VectorautoregressiveVARX VectorautoregressivewithxvariablesVE ValueequityV2C Value-to-customer

V2F Value-to-firm

1Bigdatachallenges

Introduction

Oneofthebiggestchallengesfortoday’smanagementliesintheincreasingprevalenceofdata. This is frequently referred to as “big data”. A recent study by IBM among chiefmarketing officers (CMOs) indeed reports that big data or the explosion of data isconsideredamajorbusinesschallenge(IBM,2012).Oneofthemainunderlyingdriversofthisexplosion is the increasingdigitalizationofoursociety,businessandmarketing.Onecan hardly imagine that consumers around the globe nowadays could live withoutsmartphones, tablets, Facebook and Twitter. Marketing is probably one of the businessdisciplines most affected by new developments in technology. In the last decades,technologicaldevelopmentssuchasincreasingdata-storageinsteadofdata-storecapacity,increasing analytical capacity, increasing online usage, etc. have dramatically changedaspects of marketing. More specifically, we have seen the development of customerrelationshipmanagement,orCRM(Kumar&Reinartz, 2005).ThisarrivalofCRMposedchallenges formarketing and raised issues on how to analyze and use all the availablecustomerdatatocreate loyalandvaluablecustomers(Verhoef&Lemon,2013).Withthegenerationofevenmoredataandothertypesofdata,suchastextandunstructureddata,firmsconsiderhowtousesuchdataasanevenmoreimportantproblem.ArecentstudybyLeeflangandVerhoefinjointcooperationwithMcKinseyconfirmsthis(Leeflang,Verhoef,Dahlström,&Freundt,2014).Theyfindthatmarketingisstrugglingwithgainingcustomerinsightsfromtheincreasingamountofavailabledata.AccordingtoMcKinsey,oneofthemainexplanations is a lackofknowledgeand skillsonhow toanalyzedataandhow tocreatevaluefromthesedata.

Explosionofdata1

Datahavebeen around for decades.However, thirty to fortyyears ago, these datawereusually available at an aggregate level, such as a yearly or monthly level. Withdevelopmentssuchasscanningtechnologies,weeklydatabecamethenorm.Inthe1990s,firmsstartedtoinvest in largecustomerdatabases,resultinginthecreationofrecordsofmillionsofcustomersinwhichinformationonpurchasebehavior,marketingcontacts,andothercustomercharacteristicswerestored(Rigby,Reichheld,&Schefter,2002).ThearrivaloftheInternetandmorerecentlyofsocialmediahaveledtoafurtherexplosionofdata,anddailyorevenreal-timedatahavebecomeavailabletomanyfirms.It isbelievedthatgetting value from these data is an important growth engine and will be of value toeconomiesinthecomingyears(seeFigure1.1).

Figure1.1EffectsofnewdevelopmentsincludingbigdataonGDP

Source:FigureadaptedfromMcKinseyGlobalInstitute(2013)

The Internet has become one of themost important marketplaces for transactions ofgoods and services. For example, online consumer spending in theUnitedStates alreadysurpassed $100 billion in 2007, and the growth rates of online demand for informationgoods, such as books, magazines, and software, are between 25 and 50 percent(Albuquerque,Pavlidis,Chatow,Chen,& Jamal, 2012). In theUnitedStatesdigitalmusicsalesin2011exceededphysicalsalesforthefirsttimeinhistory(Fisch,2013).BesidesB2CandB2Bmarkets,onlineC2Cmarketshavegrownin importance,withexamplessuchasLuLu,eBayandYouTube.Thenumberof Internetusersbytheendof2014wasover279millionintheUnitedStatesandmorethan640millioninChina(InternetLiveStats,2014).Worldwide, there are about 1.4 billion active users of Facebook at the end of the firstquarterof2015.OnaverageTwitterusersfollowfivebrands(Ali,2015).Companiesarealsoincreasingly investing in social media, indicated by worldwide marketing spending onsocialnetworkingsitesofabout$4.3billion(Williamson,2011).Managersinvestinsocialmedia tocreatebrandfans,as this tends tohavepositiveeffectson firmwordofmouth

and loyalty (Uptal & Durham, 2010; De Vries, Gensler, & Leeflang, 2012). There are 32billionsearchesonGoogleeverymonthand50millionTweetsperday.Theuseofsocialmediaalsocreatesa tremendous increase incustomer insights, includinghowconsumersareinteractingwitheachotherandtheproductsandservicestheyconsume.Blogs,productreviews, discussion groups, product ratings, etc. are all new important sources ofinformation (Onishi & Manchanda, 2012; Mayzlin & Yoganarasimhan, 2012). Theincreasing use of online media, including mobile phones, also allows firms to followcustomersintheircustomerjourneys(Lemke,Clark,&Wilson,2011).

Bigdatabecomethenorm,but…

If one considers the popular press, big datahavenowbecome thenormand firmshavestarted to understand that theymight be able to competemore effectively by analyzingthese data (e.g. Davenport&Harris, 2007). There are several popular examples of firmsanalyzingthesedata,suchasIBM,Tesco,CapitalOne,Amazon,Google,andNetflix.Butmany companies struggle with getting value from these data. Besides, firms can easilybecomedisappointedabout theirefforts regardingbigdataanalytics,aswehave seen inearlierdata revolutions, suchasCRM(e.g.Verhoef&Langerak,2002).Oneproblemwasthe dominant role of IT inCRM implementation. The samemay happenwith big data.Moreover,bigdatadevelopmentshavestirredupvigorousdiscussionandpublicconcernonprivacyissues.ThesediscussionsandconcernshavebecomeevenmoreprevalentasaconsequenceoftheactionsofEdwardSnowden,wholeakeddocumentsthatuncoveredtheexistenceofnumerousglobalsurveillanceprograms,manyofthemrunbytheNSAandtheFive Eyes with the cooperation of telecommunication companies and Europeangovernments.2Butstillfirmsunderestimatetheprivacyreactionsofcustomersandsocietalorganizations. For example,when theDutch-based bank INGannounced that theyweregoing to use payment information to provide customers with personalized offers andadvice, strong reactionson (social)media arose and even theCEOof theDutchCentralBanksaidthatbanksshouldbeveryhesitantwiththiskindofbigdatainitiative.

Theproblemswithcreatingvaluefrombigdatamainlyariseduetoalackofknowledgeand skills on how to analyze and use these big customer data. In addition, firmsmightoverestimatethebenefitsofbigdata(Meer,2013).Oneimportantdangeristhatfirmsstarttoooptimisticallyandstartthinking“toobig”,whileactuallylackingdecentknowledgeonthebasicsandchallengesofgooddataanalysisofalreadyexistingdata,suchasCRMandsurvey data, andhow this can contribute to business performance. Firms start up large-scalebigdataprojectswithratherdifficultdataminingandcomputersciencetechniquesandsoftwareprograms,withoutaproperdefinitionoftheobjectivesoftheseprojectsandtheunderlyingstatisticaltechniques.Asaconsequence,firmsinvestheavilyinbigdatabutarelikelytofaceanegativereturnoftheirbigdatainvestments.

Ourobjectives

Given the growing importance of big data, their economic potential, and the problemsfirms face on capitalizing on these opportunities, we believe there is an urgent need toprovidemanagerswithguidanceonhowtosetupbigdatainitiatives.Bywritingthisbookwe aim to providemanagerswith this guidance. Specifically themain objectives of thisbookarethreefold:

Our first objective is to teachmanagershow the increasingpresenceofnewandlargedataprovidesnewopportunities tocreatevalue.For thatreason,wediscussnotonlythe increasingpresenceof thesedata,butalso importantvalueconcepts.However, we also consider the possible dark sides of big data and specificallyprivacyanddatasecurityissues.As a second objective, we aim to show how specific analytical approaches arerequired, how value can be extracted from these data and new growthopportunitiesamongnewandexistingcustomersdeveloped.Thirdly,we discuss organizational solutions on how to develop and organize themarketinganalyticalfunctionwithinfirmstocreatevaluefrombigdata.

Ourapproach

Althoughwebelieveinthepotentialpowerofanalyticsandbigdata,weaimtoprovideamorenuancedviewonbigdatadevelopments.Inessence,webelievethattheexistenceofbigdatainitselfisnotarevolution,itisratheranevolutionoftheincreasingavailabilityofdata observed in recent decades as a result of scanner data developments, CRM datadevelopmentsandonlinedatadevelopments.Bigdataaremakingdatadevelopmentmoremassiveandthisalsoleadstonewdatasources.Despitethis,manyanalyticalapproachesremainsimilarandknowledgeon,forexample,howcustomerandmarketingintelligenceunits have developed, remains valuable. Building on extensive academic and practicalknowledgeonmultipleissuessurroundinganalytics,wehavewrittenabookthataimstoprovide managers and analysts with strategic directions, practical data and analyticalsolutionsonhowtocreatevaluefromexistingandnewbigdata.Todoso,thisbookhastwo specific approaches. First, we aimed to write a book that is useful for marketingdecisions onmultiple levels. Typically there has been a kind of disconnect between, forexample,brandmanagementandcustomermanagement(Leoneetal.,2006). Inthisbookwediscusstheuseofbigdataatthreelevels:

1. marketlevel;2. brand/productlevel;and3. customerlevel.

Wetakethisapproachbecauseweobservethatbigdatahaveanimpactonalltheselevels.Typicalbrand-orientedfirms,suchasUnileverandPhillips,areasinterestedinbigdataasfirmswith individualcustomer leveldata, suchas INGandAmazon.Moreover,bigdataprovideopportunitiesfordataintegrationandinsightsusingdatafrommultiplelevels.

Second,wehaveauniquecombinationofascientificandpracticalapproachtobigdataandcustomeranalytics.Withinmarketingsciencewehaveobservedincreasingattentiontocustomer andmarketing analytics (Verhoef, Reinartz, & Krafft, 2010; Verhoef & Lemon,2013), which has provided extensive knowledge on theoretical CRM concepts such ascustomer lifetime value (CLV). Furthermore, specific models have been developed, forexampletopredictcustomerloyaltyandvalue(e.g,Neslin,Gupta,Kamakura,Lu,&Mason,2006; Venkatesan & Kumar, 2004). However, despite this increasing presence, marketingscienceandanalyticalpracticearefrequentlyseparated.Usingourknowledgefromscienceand practice, we aim to provide a scientifically solid, pragmatic and usable approachtowardscreatingvaluefromdatawithinfirms.Wewillprovideanumberofcaseswithineach chapter to show how our discussed concepts and techniques can be used withinmarketingpractice.Weuseanovelapproachinthewaythisbookisdividedintochapters.The main chapters present an overarching discussion on the main theoretical andconceptual ideason, forexample,bigdata,valuecreationandanalytics.Beyond thatwe

havesecondaryin-depthchaptersthataimtoprovidetheinterestedreaders(e.g.thedatascientist)withmuchmorein-depthknowledgeonthesespecificconceptsandanalytics.Assuch, thisbookcanbeveryvaluable for (marketing)managersaiming tounderstand thecore concepts of big data analytics in marketing, and also for marketing and customerintelligencespecialistsanddata-scientists.

Readingguide

ThestructureofourbookisdisplayedinFigure1.2.Westartwithtwogeneralchapters(ofwhich this introduction is the first). In these chapters we discuss our main underlyingvision on big data and customer analytics and the relevance of analytics for firms. InChapter 2 we discuss our main big data value creation model that will be used as aguidance for the following chapters. Next we have key chapters which focus on thebusinessmanagement level:we focuson theomnipresenceofdata (Chapter3),analytics(Chapter4)andthedevelopmentofananalyticalorganization(Chapter5).ForChapters2,3and4wehavewrittenunderlyingin-depthchapters.Forexample,forvaluecreationwefocusonspecificmetricsofourvalueconcepts:value-to-firm(V2F)andvalue-to-customer(V2C). Similarly, in-depth chapters on analytics discuss analytical classics, big dataanalytics and story-telling and visualization. As previously mentioned, the function ofthesein-depthchaptersistoprovidereaderswithmoredetailedknowledgeand/ortoolsforeachofthemorehigh-leveltopicsdiscussedinthehigher-levelchapters.InChapter6wedescribe specific cases in (bigdata) analytics.Weendby settingout themost importantlearningpoints.

Figure1.2Readingguideforbook

Weurgethereadertostartfirstwiththegeneralandkeychapters.Thein-depthchapterscannotbereadindependentlyfromthegeneralandkeychapters!Ifonelikestohavemoredetailedknowledgeon specific topics one can laterpick and choose from these in-depthchapters.

Notes

1ThissectionisbasedonLeeflang,Verhoef,Dahlström,&Freundt(2014).

2 See https://en.wikipedia.org/wiki/Global_surveillance_disclosures_(2013%E2%80%93present) (accessed September 14,

2015).

References

Albuquerque, P., Pavlidis, P., Chatow, U., Chen, K., & Jamal, Z. (2012). Evaluatingpromotional activities in an online two-sided market of user-generated content.MarketingScience,31(3),406–432.

Ali,A.(2015).Whydowefollowbrandsonsocialmedia?RetrievedfromSocialMediaToday.RetrievedJune10,2015from:www.socialmediatoday.com/social-business/asadali/2015-05-24/business-social-media-infographic.

Davenport,T.,&Harris, J. (2007).Competingonanalytics–Thenewscienceofwinning.HarvardBusinessSchoolPress.

DeVries,L.,Gensler,S.,&Leeflang,P.S.H.(2012).Popularityofbrandpostsonbrandfanpages:Aninvestigationoftheeffectsofsocialmediamarketing.JournalofInteractiveMarketing,26(2),83–91.

Fisch,K.(2013).Didyouknow3.0.RetrievedJanuary19,2013from:www.youtube.com/watch?v=jp_oyHY5bug.

IBM.(2012).Analytics:Thereal-worlduseofbigdata–Howinnovativeenterprisesextractvaluefromuncertaindata.IBMInstituteforBusinessValue.RetrievedSeptember11,2015fromwww.ibm.com/smarterplanet/global/files/se__sv_se__intelligence__Analytics_-_The_real-world_use_of_big_data.pdf

InternetLiveStats.(2014)InternetUsersbyCountry.RetrievedfromInternetLiveStats.RetrievedJune10,2015fromwww.internetlivestats.com/internet-users-by-country/.

Kumar, V., & Reinartz, W. (2005). Customer Relationship Management: A DatabasedApproach.USA:JohnWileyandSons.

Leeflang, P. S. H., Verhoef, P. C., Dahlström, P., & Freundt, T. (2014). Challenges andsolutionsformarketinginadigitalera.EuropeanManagementJournal,32(1),1–12.

Lemke,F.,Clark,M.,&Wilson,H.(2011).Customerexperiencequality:Anexplorationinbusiness and consumer contexts using repertory grid technique. Journal of theAcademyofMarketingScience,3(6),846–869.

Leone, R. P., Rao, V. R., Keller, K. L., Luo, A.M.,McAlister, L., & Srivastava, R. (2006).Linkingbrandequitytocustomerequity.JournalofServiceResearch,9(2),125–138.

Mayzlin,D.,&Yoganarasimhan,H.(2012).Linktosuccess:Howblogsbuildanaudiencebypromotingrivals.ManagementScience,58(9),1651–1668.

McKinseyGlobalInstitute.(2013).Gamechangers:FiveopportunitiesforUSgrowthandrenewal.RetrievedfromMcKinsey.com.RetrievedSeptember11,2015fromwww.mckinsey.com/insights/americas/us_game_changers.

Meer,D.(2013).TheABCsofanalytics.StrategyBusiness,70,6–8.

Neslin, S. A., Gupta S., Kamakura, W. A., Lu, J. X., & Mason, C. H. (2006). Defectiondetection: Measuring and understanding the predictive accuracy of customer churnmodels.JournalofMarketingResearch,43(2),204–211.

Onishi,H.,&Manchanda,P. (2012).Marketingactivity,bloggingandsales. InternationalJournalofResearchinMarketing,2(3),221–234.

Rigby,D.K.,Reichheld,F.F.,&Schefter,P.(2002).AvoidthefourperilsofCRM.HarvardBusinessReview,82(11),101–109.

Uptal, M. D., & Durham, E. (2010). One cafe chain’s Facebook experiment. HarvardBusinessReview,88(3),26–26.

Venkatesan, R., & Kumar, V. (2004). A customer lifetime value framework for customerselectionandresourceallocationstrategy.JournalofMarketing,68(4),106–215.

Verhoef,P.C.,&Langerak,F. (2002).Elevenmisconceptionsabout customer relationshipmanagement.BusinessStrategyReview,13(4),70–76.

Verhoef,P.C.,&Lemon,K.N.(2013).Successfulcustomervaluemanagement:Keylessonsandemergingtrends.EuropeanManagementJournal,31(1),1–15.

Verhoef, P. C., Reinartz, W. J., & Krafft, M. (2010). Customer engagement as a newperspectiveincustomermanagement.JournalofServiceResearch,13(3),247–252.

Williamson,D.A.(2011).Worldwidesocialnetworkadspending:Arisingtide.RetrievedfromeMarketer.com.RetrievedSeptember,2015fromwww.emarketer.com/Report.aspx?code=emarketer_2000692.

2Creatingvalueusingbigdataanalytics

Introduction

Nowadays, the existence of big data is such a hype that firms are investing in big datasolutionsandorganizationalunitstoanalyzethesedataandlearnfromthem.Weobservethatfirmsarenow,forinstance,hiringbigdatascientists.Thisoccursinallsectorsoftheeconomyincludingtelecom,(online)retailing,andfinancialservices.Firmshaveastrongbelief that analyzing big data can lead to a competitive advantage and can create newbusinessopportunities.

However, at the same time experts are warning of too high expectations. Somecommentators even consider big data as being only a hype that will mainly providedisappointingresults.1DavidMeer (2013) suggests that takingahistoricalperspectiveonearlierdataexplosionsshowsspecificpatternsinthebeliefsaboutthepotentialbenefits.Hespecificallyrefers to thescanningrevolution in the1980sand theCRMrevolution in thelate1990s(Verhoef&Langerak,2002).Firmstypicallygothroughthreestages:

1. Dataenthusiasm—Investmentphase2. Datadisappointment—Frustrationdisinvestmentphase3. Datarealism—Reinvestmentphase

Inthefirstphasetherearestrongbeliefswithinafirmaboutthepotentialbenefitsthatcanbe achieved. Frequently, top management is seduced by enthusiastic examples in thebusinesspressandeffective sales strategiesof IT,managementconsultants, and softwareproviders.However,aftersomeyearsthedataexplosioninvestmentsandinitiativesprovidemainly disappointing results and failed projects occur frequently. This induces firms torethink their data strategies and sometimes disinvest in data initiatives and IT. Thisrethinking of strategies is usually the stepping stone towards a next phasewith refinedexpectations,morerealisticambitionsandastrongerfocusonthevaluecreatingpowerofdata-based initiatives and its return on investment (Verhoef & Lemon, 2013; Rigby &Ledingham,2004).

Of course firms can go through these phaseswhen implementing big data initiatives.However,thiswouldcertainlyleadtovaluedestruction,negativeROIs,wasteofresources,and enormous frustration. Instead of going through these phases,we propose that firmsshould have sound initial expectations on the value of potential big data. For this, it isessentialtounderstandhowbigdatacancreatevalue.Furthermore,itisourstrongbeliefthat firms shouldunderstand their analytical strategies and the approach they choose inanalyzingavailabledata.

Inthischapterwelayoutthefoundationsforasoundvalue-creatingbigdatastrategy.Wediscusshowbigdatacancreatevalueandwhatelementsarerequiredtocreatevalue.

Bigdatavaluecreationmodel

Oneofthebiggestchallengesofbigdataishowfirmscancreatevaluewithbigdata.Wehavedevelopedthebigdatavaluecreationmodeltoshowhowthisvaluecreationoccurs(seeFigure2.1).Thismodelhasfourelements:

1. Bigdataassets2. Bigdatacapabilities3. Bigdataanalytics4. Bigdatavalue

Bigdataassets

Figure2.1Bigdatavaluecreationmodel

Assetsareusually consideredas resourceendowments thata firmhasaccumulatedovertime. These assets can be tangible (e.g. plant) or intangible (e.g. brands, customerrelationships). Inthepast,customerdatabaseswereconsideredimportantassetsforfirms(Srivastava,Tasadduq,&Fahey,1998).Forexample,thesedatabasescouldbeusedtocreatestrongerrelationshipswithcustomers,achievehigherloyalty,andcreatemoreefficientandeffective(cross)-sellingtechniques.Inaneraofbigdata,thedataarenolongerrare.Onecouldactuallyarguethatthedataarenolongerthatvaluable,asdataareomnipresent,canbecollectedinmultiplewaysandarefrequentlypubliclyavailabletomanyfirms(e.g.dataononlinereviews).Inprinciple,westronglysympathizewiththisview.However,wealsoobserve thatwithin firms there is actually a lackof knowledgeon themerepresenceofdata within the firm itself and outside the firm. For example, one of the largest cablemanufacturing companies in Europe only recently discovered that by diving into someinternal billing data, they could gain valuable insights on loyalty and customer lifetimevalue (CLV) developments. We will discuss the different sources and types of data inChapter3.

Bigdatacapabilities

We can see that the value of data is not in the mere presence of the data, but in theunderlyingcapabilitiesabletoexploitthesedata.Weconsidercapabilitiesasthe“glue”thatenablesbigdata—simultaneouslywithotherassets—tobeexploited to createvalue (Day,1994).Forexample,usingdifferentdatasourcesoncustomerexperiences,onecouldlearnhow to improve these experiences, thereby also building on the qualitative input of keycustomers(relationalasset)thatmayfurtherimprovethecustomerexperience.

Theseunderlyingcapabilitiesthatcanbeusedonbigdataconcern:

1. People2. Systems3. Processes4. Organization.

People

To exploit big data, people are very important.Without the right set of skilled big dataexperts it is not sensible todevelop a bigdata strategy.Having intelligencedepartmentswith the right capabilities is of essential importance (Verhoef & Lemon, 2013). This isactuallyoneofthebiggestchallengesforfirms(Leeflang,Verhoef,Dahlström,&Freundt,2014).Firmsarenowhiringbigdatascientists,butthesepeoplearedifficulttofind.Asaconsequence, firmshave also chosen to educate big data scientists in-house through, forexample, specific internal programsandacademies (Verhoef&Lemon, 2013).Given thatpeople are of essential importance for a successful big data strategy, we will devote aspecial chapter to how firms can develop a strongmarketing intelligence capability (seeChapter5).

Systems

With regard to systems, we strongly emphasize the importance of data integration andproviding an integrated data ecosystemallowing the firm to analyze data frommultiplesources. We still observe that within firms data are collected in different systems ordatabases, which are not sufficiently linked. This data integration requires specific datamanagementskillsandsoftware.Dataintegrationbecomesevenmoredifficultwhenfirmsare operating inmultiple channels or inmultiple countries where different systems arebeingused(Neslinetal.,2006).Akeyquestionforfirmsistowhatextentdatashouldbeintegrated,asthemarginalreturnsondataintegrationmightdecline(Neslinetal.,2006).Animportanttrendwithsystemsisthat,duetothesizeofbigdata,(cloud)solutionssuch

as Hadoop have been developed. Similarly, we observe several new trends in availableanalyticalsoftware.Oneofthemajortrendsisthedevelopmentofopensource“packages,”suchasR,whichcanbeused for free.Although this involvesa lotofprogramming, theprogramsarewidelysharedbetweencommunitiesofusers,sothatthesepackagesbecomemoreeasilyaccessible.Wewillhaveamorein-depthdiscussiononsystemsandspecificallydata-basedsolutionsandsoftwaresolutionsinChapter5.

Process

Processeswithregardtosmartbigdataanalyticsmainlyconcernhowfirmsorganizethedatainputandstorage,theaccessibilityofdatatoanalyticalteamsandthecommunicationbetweenanalyticteamsand(marketing)management.Thefirsttwoprocessesarerelevantforsmoothandreal-timedataaccessibility.Importantly,theseprocessesalsoinvolvehowfirmsdealwith privacy, data security issues, and legal issueswith regard to data usage.Privacyand securityhavebecomea toppriority for firmsandboth receive considerableattentionamongpolicymakersasaresponsetotheincreasingavailabilityofbigdataandscandalsinvolvingbigdata.Thetrendseemstobethatlegislatorsarereducingthefreedomof firms to use individual customer-level data. As a consequence, firms are becomingstricterwithdatausageandstorage.Forexample,weknowoffirmsthatstoredcustomerdata covering several years, but now only store transaction data of customers for amaximum period of a year. Data security is becoming an issue: there have beenmanyexamples of hackers and criminal organizations being able to illegally get data on, forexample, passwords, payment data (e.g. credit card numbers) and other personal data.Hackers are not the only problem—employees who are less careful with data (e.g. loselaptops or throwawaydata storage deviceswith sensitive data on them) can also causesecurityproblems.Datacomplianceisthusanimportantelementofbigdataprocesses.Theusageofthesedatacanhurtmillionsofcustomersaroundtheglobe.Theotherpartoftheprocessesconcernshowmarketingandanalyticalteamscommunicate.Thisinvolvesatwo-way communication. On the one hand marketing should clearly communicate tomanagementtheproblemsandchallengestheyfaceandhowanalyticscouldbehelpfulinsolving them. On the other hand analytical teams should be able to effectivelycommunicate their findings through insightful reports and marketing dashboards.Moreover,inanerawherebigdataanalyticscancreatevalue,analyticalteamsshouldbeable to effectively communicate big data-based value-creating solutions to themanagement.Theseprocesseswillprobablydevelopinanaturalway,butitmightalsobeimportanttodefineprocessesupfrontinwhich,forexample,marketingisrequiredtogetintouchwiththeiranalyticalteamswhenamarketingproblem(e.g.adecreaseinloyalty)is observed. Processes on how marketing dashboards should be fuelled with relevantinformationovertimeshouldalsobedefined.

Organization

Beyond having good people, firms also need to devote attention to how big data andspecifically big data analytics can be organized internally. One crucial question in thisrespectiswhetheranalyticsorintelligencedepartmentscanreallyhaveanimpactondailybusiness.Weobservedseveralmodelsonhowtheanalyticalfunctionisembeddedwithinfirms. Typically, intelligence functions are separate staff departments that serve themarketingand sales functionswithoutcomesof theiranalyses, eitheron requestor self-initiated.However,inordertohaveastrongerimpact,somefirmschoosetointegratetheintelligencedepartmentwiththemarketing/salesdepartment.Theunderlyingideais thatthiswillinduceastrongeruseofanalyticswithinmarketingdecisionmaking(Hagenetal.,2013).Morelikely,however,theresultisareductionintheindependenceoftheanalyticsdepartment,withnegativeconsequences,suchasalackofinnovationandnotsufficientlythought-through analyses. A disadvantage of such an organization might also be thatanalyticalknowledgeisnotusedoptimallywithintheorganizationasitisfragmentedovermultipledepartmentsand/orfunctions.

Theroleofculture

One of themost prevalent issues in exploiting big data as an asset is the nature of theinternalcultureandtherelatedprocesses.Traditionally,marketinghasbeenafunctionthattended to rely on intuition and gut feeling. Fortunately, only having a good idea is nolongergoodenough inmanyfirms(DeSwaanArons,VandenDriest,&Weed,2014). Infactthereisanincreasingtrendtowardsmoredata-drivenorfact-baseddecisionmaking,partially explained by a stronger emphasis on marketing accountability (Verhoef &Leeflang,2009).Bigdataanalyticscanonlysurvivewithin firms thatembrace this trendandindeedareopentorelymoreonanalyticsandtheirresultinginsightsandmodelsthatprovide ideas for innovation,orshowtheeffectivenessofspecificmarketingactions,etc.This requires a strong move within firms and specifically marketing departments. Thischange in culture can be rather dramatic. Old-school marketers have to change theirdecision-makingstyleandhavetogainmoreknowledgeonanalyticsandhowtheycanbeusedtomakesmartermarketingdecisions.Thisrequiresintensiveeducationprogramsfor—or inextremecases replacementof—thesemarketers.One specific challenge, though, ishowtheanalyticalleft-brainculturecanbecombinedwithamorecreative/intuitiveright-brain culture (Leeflang et al., 2014; De SwaanArons et al., 2014). In Chapter 5wewilldiscusstheissuessurroundingbigdatacapabilities.

Bigdataanalytics

Reading a book about big data and analytics, onewould probably expect that analyticswoulddeserveimmediateattention.However,analyticsnotembeddedintheorganizationwithouttherelevantdata,culture,andsystemswillhavelimitedimpactandvalue-creatingpotential.Whendiscussingbigdataanalytics,wemakeadistinctionbetweentwodifferentformsofanalytics:

AnalyticsfocusingongaininginsightsAnalyticsaimingtodevelopmodelstoimprovedecisionmaking.

Wedefinebigdatainsightsusuallyasdescriptivefindingsresultingfromdataanalysesthatprovide input into marketing decisions. Models are purposely developed to direct andsupport marketing decisions. Model development is almost like an R&D task in whichanalystswork to an end goal on amodel,which is accepted by themanagement of thedepartmentandusersofthemodels(e.g.VanBruggen&Wierenga,2010).

Thedevelopedinsightsandmodelscancreatevalueforfirmsinthreeways:

DecisionsupportformarketingImprovedactionsandcampaignsInformation-basedproductsandsolutions

Using the developed insights and models firms can potentially make more informeddecisionsonwhere toallocate theirmarketingbudgets.Resultsofamodelcanshowthespecificeffectivenessofanadvertisingchannel.Forexample,whenDeVries(2015)showedthe limited influence of socialmedia on acquisition, one could questionwhether a firmshould heavily focus on social media to attract new customers. Leeflang et al. (2014)distinguish between two different models that can be developed to drive marketingdecisionmaking:

Idiosyncratic, usually more sophisticated models developed to tackle specificmarketingproblemsStandardizedmodels thathavebecome important tools to improve thequality oftacticalmarketingdecisions.

Themarketingliteraturehasidentifiedmanystandardizedmodels(e.g.ScanPro),whicharemainly delivered bymarketing research agencies such as ACNielsen, IRI and ResearchInternational (Hanssens, Leeflang, & Wittink, 2005). These standardized models can befilled with available data within firms and research agencies. We expect that researchagencieswill providemore standardized solutions on howbig data can be integrated togaincustomerinsightsandestimatetherelationshipsbetweenmarketinginstrumentsand

marketingoutcomes.

TheimprovementofactionsandcampaignsismainlyrelevantinaCRMenvironment.Itmainlyhastodowithwhomtotarget,whentotargetandwithwhatmessage.Ithasbeenshown that through effective selection of customers, the ROI of campaigns can beimproved (e.g. Bult & Wansbeek, 1995). It has been observed that customization ofmessagesandoffers,specificallyinanonlineenvironment,canbeveryvaluable(Ansari&Mela,2003).Inabigdataenvironmentthisnowoccursinreal-time,andisalsoknownasbehavioraltargeting.Thiscan,however,havenegativesideeffectsasitmaybeconsideredintrusive(VanDoorn&Hoekstra,2013).

Arelativelynewdevelopmentintheeraofbigdataistheuseofresultsofanalysesandmodels to develop information-based products and solutions that specifically focus oncustomers to createvalue for these customers. For example, anovelplayer in theDutchbankingsector,KNABbank,isexplicitlyprovidingdata-basedsolutionstotheircustomersto advise themhow to use their availablemoney (e.g. put it in a savings account). TheDutchrailwaysprovideaservicetotheircustomersinwhich,basedonactualinformationon traffic and trains, the fastest transportmode is recommended (Leeflang et al., 2014).NobelprizewinnerRichThalerbelievesthatthesesolutions,eitherdevelopedbysuppliersthemselvesor byother, frequently independent, infomediaries,will become important inhelpingcustomerstomakemoreinformeddecisions(Thaler&Tucker,2013).

Strategiesforanalyzingbigdata

The presence of big data provides huge opportunities for analytical teams. One of theeasiest ways of using it is probably just to start up analyses and start digging into theavailabledata.Bydigginginthedata,onemightgainveryinterestinginsights,whichcanguide marketing decisions. The most famous example in this respect is the UK-basedretailerTesco:whenanalyzingdataof their loyaltycard, theydiscoveredthatconsumersbuying diapers also frequently buy beer and chips (Humby, Hunt & Phillips, 2008).Althoughsuchanexamplecanbeinspiring,wepositthatbeforestartingupananalyticalexercise, one should clearly understand the benefits and disadvantages of this specificanalysis strategyaswellas thatofother strategies.Thereforewestronglyadviseamoreproblem-driven approach instead of a rather exploratory findings approach.We discussthesestrategiesinmoredepthinChapter4.

Bigdataischanginganalytics

Bigdataisbelievedtochangeanalyticsasbigdatahasspecificcharacteristicsknownasthe3Vs of big data, posing specific challenges for researchers andmanagers (Taylor,Cowls,Schroeder,&Mayer,2014;Leeflangetal.,2014):

IncreasingdataVolumeIncreasingdataVelocityIncreasingdataVariety

Theincreasingvolumeofdataimpliesthatdatabasesbecomeverylarge,andtheanalysisofdataofmillionsofcustomerswithhundredsofcharacteristicsisnolongeranexception(e.g.Reimer,Rutz,&Pauwels, 2014).Dataarealsoarrivingmorequickly,which inducesfasteranalysisandfasteraction(Leeflangetal.,2014).Wehavebeenmovingfromyearlydatatomonthlydata,toweeklydata,todailydataandnoweventodataperhour/minute.Finally,thedataarebecomingmorecomplexastheyarriveindifferentformats.Inthepastnumerical data was the standard. Nowadays, more unstructured data such as text andaudiodata are also available, and also videodata through, for example,YouTube.OtherexamplesincludedataonFacebookpostings,andGPSdatafrommobiledevices.ThethreeVs have been extended to fiveVs,whereVeracity andValue have been added.Veracityreferstothemessinessandtrustworthinessofdata.Withtheincreasingavailabilityofdata,notalldataareasreliableasonewouldlike.Hence,dataqualitycanbelow.Forexample,itisknownthatcustomerreviewsarebeingmanipulated.Valueisconsideredasthevaluethatiscapturedfromanalyzingandusingthedata.Althoughweclearlydoacknowledgethatvalue shouldbecaptured (seeourbigdatavaluecreationmodel) it isnota specificcharacteristicofbigdata,whichischanginganalytics.

Howthesebigdataarechangingmarketinganalyticsisnotasclear.Marketingscientistshave argued the following: high volume of data implies the need for models that arescalable;highvelocityopensopportunities for real-time,orvirtually real-timemarketingdecision making that may or may not be automated; and high variety may requireintegration across disciplineswith the corresponding sensitivity to variousmethods andphilosophiesofresearch.2Insum,thissuggeststhatmodelsshouldeasilybeestimatedonlarge sample sizes,whereas analytics should be done in such away that it can provideimmediateresults,andfinallynewmethodologiesfromotherdisciplines,suchascomputerscienceandlinguistics,shouldbeintegrated.

However,wealsowarnanalystsandmanagersthatdespitethedifferentcharacteristicsofbigdatacomparedtotraditionaldata,oneshouldalsobecarefulofimmediatelymovinginto a totally different analysis mode. For example, despite the huge volume of dataavailable, analysis can still be done on smaller samples of the available data. Theinformationpresentinunstructureddatamayalsobemorelimitedthanexpected.DeVries(2015) recently showed that the additional explanatory power of Facebook “likes” inexplainingsalesisrather limited. It isourcontentionthat inorder tobeagoodbigdataanalyst and tousebigdata in a good fashion, one shouldmaster thebasics of analyticsratherthanmovingimmediatelyintograndbigdataanalyticalexerciseswithoutactuallyknowingwhatoneisdoing.

Thepowerofvisualization

Analysts with a left-brain who are trained in statistics will find it easy to understandnumerical outcomes of analyses. However, for many other people understanding themeaning of numbers is quite difficult. Presentation of analyses (and their results) istherefore a crucial task when analyzing the data. One way to have more impact is tovisualize the data (that is, using visual aids in data presentation), because humans ingenerallookforstructures,anomalies,trends,andrelationships.Visualizationsupportsthisby presenting the data in various forms with different interactions. It can provide apowerfulqualitativeoverviewofdataandanalyticalresults.Itcanalsoshowtheimportantrelationshipsinthedata(Grinstein&Ward,2002).Webelievethatvisualizationisaveryimportantanalyticalcapabilitywhoseimportanceisfrequentlyneglected.Itdoes,however,allow researchers to havemore impact on dailymarketing decisions, as it enhances theaccessibility of analytical results for especially right-brain trainedmarketing executives.Despitethis,oneshouldalsobeverycareful.Visualizationcanleadtoanoversimplificationof results (e.g. by providing a scatter plot of a spurious correlation) or can easilyoverestimate the found effects with some scaling tricks on graphical axes. Hence, oneshouldalsobecarefulnottocommunicateastatisticalillusionwhenvisualizingthedata.

Frombigdataanalyticstovaluecreation

Weconsiderthreemethodsbywhichbigdataanalyticscancreatevalueforcustomersandfirms.First,bigdataanalyticscancreate importantnewinsights that improvemarketingdecision making. For example, big data analytics can show how firms can improvecustomersatisfaction through improving, forexample, thespecific featuresof theserviceexperience.Byhavingtheseinsightsmarketingbudgetscanbeallocatedmoreeffectively.Insteadof relyingon intuition,brandmanagers can, for example, invest inapositioningstrategythateffectivelydifferentiatesbrandsfromcompetitors.

A second value-creation benefit of big data analytics is the development of moreeffectivemarketingcampaigns,andmorespecificallymoreeffectivetargetingofcampaignsbyselectingtherightcustomers.Whereearlyanalyticsweremainlyfocusedonimmediateresponse to campaigns (e.g. Feld, Frenzen, Krafft, Peters, & Verhoef, 2013; Bult &Wansbeek,1995),alonger-termfocusisnowstronglyadvocated,achievedbyconsideringthe impact of marketing campaigns on CLV and customer equity (e.g., Venkatesan &Kumar, 2004; Rust, Lemon,&Zeithaml, 2004). The effectiveness of both approaches hasbeenshownextensivelyinthescientificliterature.Importantly,theseapproacheshavealsobeen applied in business and have been shown to increase firm value (Kumar & Shah,2009). Another development is that, especially in an online environment, real-timebehavioraltargetingisbeingusedtoadaptonlineenvironmentsandadvertisingtospecificconsideredneedsofthecustomer.

A third value-creation benefit is the development of big-data-based solutions forcustomers.Thesesolutionsdirectlyhaveanimpactoncustomersandshouldcreatemorevalue for them. Frequently, this involves an improvement of the service experience inseveralstagesofthepurchaseprocess.Forexample,specifictoolscanbedevelopedtohelpcustomersmake better purchase decisions using smart algorithms (e.g. Thaler&Tucker,2013).

Valuecreationconcepts

Valuecreationshouldbetheultimateobjectiveofeverybigdatastrategy.However,valuecreation is one of those terms that is easily written downwithout a full and completeunderstandingofthetopic.Importantly,weconsidervaluefromtwoperspectives:

1. Valuetothecustomer(V2C)2. Valuetothefirm(V2F)

These two perspectives are not novel. In fact the classical definitions of marketing putforward in basic marketing text books (e.g., Kotler & Armstrong, 2014) emphasize that

marketing should focus on creating superior value for customers (through high quality,attractivebrandpropositionsandstriving foranappropriaterelationship),and that firmscancapturevaluefromcustomersinreturnforthisvaluecreation.Thisissometimesalsoreferredtoas“valuedelivery”and“valueextraction.”Valueextractionfromcustomers isconsidered tobe adirect consequenceof valuedelivery.Value extractionoccursbypaidprice premiums, higher loyalty rates (lower churn), higher revenues per customers andstronger customer advocacy (Reichheld, 1996; Srivastasva et al., 1998). Scientific researchindeed suggests that firmswhich providemore value to customers they tend to have astrongerfinancialperformance(Anderson,Fornell,&Mazvancheryl,2004).

BalancebetweenV2FandV2C

Firmscanbeclassifiedontwovaluedimensions(seeFigure2.2).Ahighvaluedeliveryandhighvalueextractionstrategyisconsideredasawin-winstrategy.Itisusuallyseenasthebeststrategyforfirms.Despitethis,wefrequentlyobservethatfirmstendtooutperformon a single value dimension (upper left and bottom right cells). This can have dramaticconsequences.Frequently,firmstendtofocusonvalueextractionsolely:examplescanbefound in many sectors. A dramatic example is the banking industry. There has been astrongfocusonshareholders’valuewithinbanks,inducingthemtofocuslessoncustomersandthedeliveryofvaluetocustomers.Thecrisisin2008,withmanybanksfacingdifficultproblems,showedthatthissolefocusonvalueextractioncanhavesevereconsequencesforfirmsandsociety(Verhoef,2012).FirmsinotherindustriessometimesfocussolelyonV2F,totheirdetriment.Forexample,themainfocusoftheCEOoftheDutchTelco-incumbentKPNwas on creating value for shareholders.As a consequence,marketingmanagementhadastrongfocusonCLVcreationthroughcommunicationtacticsandcontractualoffers(i.e. moving from one-year to two-year contracts, minute rounding or other short-termpricing tactics). Data analysis of their CRM database, looking for potential churncandidates, was an important element of that strategy. In terms of value delivery toshareholders this strategywas rewarding: this is reflected in the top100position for theformer Dutch Telco CEO in a recent Harvard Business Review list of CEOs that weresuccessful in creating shareholder value (Hansen, Ibarra, Peyer & Von Bernuth, 2013).Despite this, after he stepped down his strategy was criticised for putting too littleemphasis on the customer, resulting in the delivery of insufficient value to customers.3

Specifically, the presumed lack of investments in service quality and innovation wasconsideredasaweaknessofthatstrategy.Infact,oneofthefirststrategiesofthenewCEOwastoannouncethatDutchTelcowouldinvestmoreinserviceandthatexcellentservicedelivery to customers would be a key strategic focus. The practice of Dutch Telco wasrather typical of many Telco companies around the globe. Actually, many firms werecriticisedforpoorserviceandtakingadvantageofcustomersthrough,forexample,usingcomplex pricing plans. In the US, Virgin Mobile was one of the firms to address this

criticism,bydevelopingcustomerfriendlypricingplansandfocusingonasuperiorservice(McGovern,2007).

Figure2.2Value-to-customervs.value-to-firm

Source:AdaptedfromReinartz(2011),andWieseletal.(2011)

Amismatchbetweendeliveredcustomervalueandextractedfirmvaluecanalsooccur(upperleftcell).Thesefirmsareratherattractiveforcustomers,butthefirmsfailtoextractmore value through, for example, achieved higher loyalty and higher price premiums.Manystartingonlinefirmsstrugglehere;theyprovidemuchvalueintermsoffreeservices,and lower prices, etc. but find it difficult to keep customers and/or ask for fees for theservicestheyprovide.Empiricalresearch,however,showsthatwhilefirmsarefrequentlyin thedownward “Enjoywhile it lasts cell,” thenumberof firms in theupper-left “Fatalattraction” cell is rather limited (Boumaet al., 2010).Wedo, however, find anumber ofexamples of firms in the “Doomed-to-fail” cell,where firmsprovide low customer valueandareunabletoextractsufficientvalue.Firmsinthiscellareinadangerousposition,astheirvaluepropositionanddeliveryrequiresstronginvestment,whileduetotheirinabilitytoextractvaluetheymightlacklong-termresourcesfortheseinvestments.ConsiderNokia:withadecreasingattractiveness forcustomersanda lowrepeatpurchase rate itnownolonger exists as an independent firm, but has been acquired in 2013 by a firm that hadsufficient financial resources to invest in theNokia brand and its underlying products—Microsoft.

V2S:Extendingvaluecreation

The above value concepts are sometimes extended. Especially in an erawhere firms areconsidered to also have a societal function and an increased focus on, for example,corporate social responsibility, a focusononlyV2CandV2F isnot sufficient (Korschun,Bhattacharya,& Swain, 2014; Porter&Kramer, 2011). In fact banks have not only beencriticised for insufficient focus on customers, but also for insufficient consideration ofsocietyasawhole(Verhoef,2012).Onecouldthereforesuggestextendingthevalueconceptbyalso taking into accountvaluedelivery to society (V2S).This couldbedone inmany

ways.Somefirms,suchasUnilever,considersustainabilityasoneoftheircoreelementsintheircorporatestrategyandaimtoshowthatintheirbusinessoperation,includingbrandpropositions. However, Procter & Gamble is using a more tactical approach and usesspecific activities at the brand level, such as dental education programs in HispanicneighbourhoodsintheUS,toshowtheirinvolvementwithlocalsocieties.

V2Scanpartlybereflected in thedeliveredvalue tocustomers. IndeedRust,ZeithamlandLemon(2000), forexample,considerbrandethicsasan integralpartof thedeliveredvalue by brands. Similarly, corporate social responsibility is considered as a driver ofcustomer satisfaction (Korschun et al., 2014) and customer satisfaction functions as animportantmediator in the effect of corporate social responsibility on customer and firmperformance (Luo&Bhattacharya, 2009;Onrust,Verhoef,VanDoorn,&Bügel, 2014). InthisbookwewillmainlytakethisperspectiveandconsiderV2SasadriverofV2C.Wealsoobserved that corporate reputation measures can be heavily correlated with customersatisfactionmetricsovertime. Inthisbookwewill thereforenotspecificallydifferentiatebetweenV2SandV2C.InsteadweconsiderV2StobeadriverofV2C.

MetricsforV2FandV2C

Metrics have become very important due to the increasing attention for accountabilitywithin firms and the resulting consequences for marketing departments (Verhoef &Leeflang, 2009). Metrics are measuring systems that quantify trends, dynamics, orcharacteristics (Farris, Bendle, Pfeifer,&Reibstein, 2010). There are numerousmetrics inmarketing that canbemeasured. Farris et al. (2010) discussmore than fiftymetrics thateveryexecutiveshouldmaster.WeclassifymetricsintoV2CandV2Fmetrics.V2Fmetricsareusuallymuchmore transaction-orientedandfocusonconcretemarketoutcomes thatcanberelatedtomonetaryconsequencesforthefirm.V2Cmetricstypicallyfocusontheevaluationofvaluebycustomers.

Figure2.3ClassificationofV2CandV2Fmetrics

BeyondthedistinctionbetweenV2CandV2Fmetrics,wedistinguishbetweenmetricsatthemarket,brand,andcustomerlevel(seeFigure2.3).V2Cmetricstypicallyfocusontheevaluationofvaluebycustomers.V2Cmetricsat themarket level include issues suchasproductawarenessandpenetrationofnewproductsandservices.BrandlevelV2Cmetricsfocus on brand evaluations and brand knowledge of customers. For example, brandawarenesswouldbea typicalV2Cmetricbut soalsoarebrandconsiderationandbrandattitudes. Some of these brand attitudes, such as brand uniqueness and brandinnovativeness,areconsideredasinputformereattitudinalbasedbrandequitymeasures.At the customer level, typicalmetrics are customer satisfaction and relationship qualitymeasures. Sometimes thesemetrics are referred toas customer feedbackmetrics (e.g.DeHaan,Verhoef,&Wiesel,2015).AverypopularV2Cmetricisthenetpromotorscore.Onemightarguethatoperationalmeasures,suchasthenumberofcomplaints,orthenumberofreported problems with the product or service, can also be considered as V2C metrics.Althoughthesemetricsaretypicallynotevaluationsofcustomersandcouldbemainlybeconsideredasinputforcustomers’perceivedvalue,theycouldbeveryvaluablemeasuresreflectingthedeliveredvaluetocustomers(e.g.,Gijsenberg,VanHeerde,&Verhoef,2015).Inthiseraofbigdata,thesemetricshavebecomemoreavailableandtheyshoulddefinitelybeconsideredinanextendedV2Cvaluecreationanalysis.

TypicalV2Fmetricsatthemarketlevelareamarketvolume,categorysales,marketsize,and number of customers. These V2Fmetrics are generally not so firm specific. At thebrand level, onewouldmeasure brand ormarket share and brand sales, and also brandequity,which is amoremonetary evaluation of a brand’s value.Ameasure that canbeused here is revenue or price premium (Ailawadi, Lehmann, & Neslin, 2003). At thecustomerlevel,CLVisacustomermetricthathasreceivedenormousattentioninthelastdecade.ItcanbeconsideredasakeyV2Fcustomermetricthatreallytriestocapturethemonetaryvaluegeneratedbyanaveragecustomeroverhisorherrelationshipwithfirms.ThismeasurecanbeextendedbyalsoconsideringCustomerEngagementValue(Kumaretal., 2010), that may include outcomes, such as referrals and actual word of mouth (e.g.Bijmolt et al., 2010). An in-depth discussion of V2C and V2F metrics can be found inChapters2.1and2.2.

Valuecreationmodelasguidanceforbook

Asnotedabove,thedifferentelementsofthebigdatavaluecreationmodelarediscussedindifferentchapters thatareyet tocome.Weextendthediagramof themodel (Figure2.1)withreferencestotherelevantchapters(seeFigure2.4),aswealsoreferredtointheabovetext.Thisfigurenicelysummarizeshowthechaptersrelatetoeachelementinourbigdatavaluecreationmodel.Furthermore,ineachchapterwestartwiththismodeltoshowwherethischapterfitswithinthismodel—thatis,guidingourdiscussiononhowbigdatacanbeused for smartmarketingdecisions. InChapter6wediscusshow firmscancreatevaluethrough an integrated approach of big data as depicted in ourmodel. Thiswill providesomeuseful applicationsonhow thebuildingblocksofbigdata analytics jointly canbeusedwithinfirms.

Conclusions

In this chapter we have discussed the big data value creation model in marketing.Understandingthismodelisessentialtounderstandingthevalue-creationpotentialofbigdataanalytics.Bigdataassetsandcapabilitiesaretheimportantbuildingblocksunderlyingbig data value creation. The capabilities involve systems, people, processes, and theorganization.Ifcapabilitiesarepresent,bigdataanalyticscanbedeployed.Thesebigdataanalytics can create marketing insights and models that subsequently can improvemarketingdecisionmaking,andimprovethesuccessofactionsandcampaigns.Moreover,itcanbeusedtodevelopinformation-basedproductsandsolutions.Weconsidervalueasamulti-dimensional construct consisting of value creation to customers (V2C) and valuecreationtofirms(V2F).TheuseofbigdataanalyticscanresultinbothmoreV2CandV2F.

Figure2.4Bigdatavaluecreationmodellinkedtochapters

Notes

1 See comment by by Nassim Nicolas Taleb that big data is “bullshit” at

www.automatiseringgids.nl/nieuws/2013/41/big-data-is-bullshit(accessedSeptember13,2015).

2http//pubsonline.informats.org/page/mksc/calls-for-papers

3 The former CEO disagrees and claims that he was not investing insufficiently in service and innovation: see

www.mt.nl/332/76174/business/ad-scheepbouwer-had-ik-dan-minder-winst-moeten-maken.html (accessed September

13,2015).

References

Ailawadi,K.L.,Lehmann,D.R,,&Neslin,S.A.(2003).Revenuepremiumasanoutcomemeasureofbrandequity.JournalofMarketing,67(4),1–17.

Anderson, E. W., Fornell, C., & Mazvancheryl, S. K. (2004). Customer satisfaction andshareholdervalue.JournalofMarketing,68(4),172–185.

Ansari,A.,&Mela,C. F. (2003). E-Customization. Journal ofMarketingResearch, 40(2),131–145.

Bijmolt,T.H.A.,Leeflang,P.S.H.,Block,F.,Eisenbeiss,M.,Hardie,B.G.S.,Lemmens,A.,&Saffert, P. (2010).Analytics for customer engagement. Journal of Service Research,13(3),341–356.

Bouma, J.T.,Bügel,M.S.,Verhoef,P.C.,Alleman,T.,Wiesel,T.,&Wesselius,T. (2010).Dutchcustomerperformanceindex:Hetnieuwemetenvanklantprestaties.TijdschriftvoorMarketing,4,58–63.

Bult, J.R.,&Wansbeek,T.J. (1995).Optimalselectionfordirectmail.MarketingScience,14(4),378–394.

Day,G. S. (1994). The capabilities ofmarket-driven organizations. Journal ofMarketing,58(4),37–52.

DeHaan,E.,Verhoef,P.C.,&Wiesel,T.(2015).Thepredictiveabilityofdifferentcustomerfeedbackmetrics for retention. International Journal ofResearch inMarketing,32(2),195–206.

De Swaan Arons, M., Van den Driest, F., & Weed, K. (2014). The ultimate marketingmachine.HarvardBusinessReview,92(7/8),54–63.

DeVries,L.(2015).ImpactofSocialMediaonConsumersandFirms,DoctoralDissertation,UniversityofGroningen.

Farris,P.W.,Bendle,N.T.,Pfeifer,P.E.,&Reibstein,D.J.(2010).Marketingmetrics:Thedefinitiveguidetomeasuringmarketingperformance.USA:PearsonEducation.

Feld,S.,Frenzen,H.,Krafft,M.,Peters,K.,&Verhoef,P.C.(2013).Theeffectsofmailingdesign characteristics on directmail campaign performance. International Journal ofResearchinMarketing,30(2),143–159.

Gijsenberg,M.J.,VanHeerde,H.J.,&Verhoef,P.C.(2015).Lossesloomlongerthangains:Modeling the impact of service crises on customer satisfaction over time. Journal ofMarketingResearch,52(5),642–656.

Grinstein,G.G.,&Ward,M.O.(2002).Introductiontodatavisualization.InU.Fayyad,G.Grinstein&A.Wierse,(Eds),InformationVisualizationinDataMiningandKnowlegde

Discovery(pp.21–46).USA:MorganKaufmannPublishers.

Hagen,C.,Khan,K.,Ciobo,M.,Miller,J.,Wall,D.,Evans,H.,&Yaday,Y.(2013).Bigdataandthecreativedestructionoftoday’sbusinessmodels.HollandManagementReview,148(4),25–37.

Hansen,M.T.,Ibarra,H.,Peyer,U.,&VonBernuth,N.(2013).Thebest-performingCEOsintheworld.HarvardBusinessReview,91(1/2),81–95

Hanssens,D.M.,Leeflang,P.S.H.,&Wittink,D.R. (2005)Market responsemodelsandmarketingpractice.AppliedStochasticModelsinBusiness&Industry,21(4/5),423–434.

Humby,C.,Hunt,T.,&Phillips,T.(2008).Scoringpoints:HowTescoiswinningcustomerloyalty.Philadelphia:KoganPagePublishers.

Korschun,D.,Bhattacharya,C.B.,&Swain, S.D. (2014).Corporate social responsibility,customer orientation, and the job performance of frontline employees. Journal ofMarketing,7(3),20–37.

Kotler,P.,&Armstrong,G.(2014).PrinciplesofMarketing.USA:PearsonEducation.

Kumar,V.,&Shah,D. (2009).Expanding theroleofmarketing:Fromcustomerequity tomarketcapitalization.JournalofMarketing,73(6),119–136.

Kumar, V., Aksoy, L., Donkers, B., Venkatesan, R., Wiesel, T., & Tillmanns, S. (2010).Undervalued or overvalued customers: Capturing total customer engagement value.JournalofServiceResearch,13(3),297–310.

Leeflang, P. S. H., Verhoef, P. C., Dahlström, P., & Freundt, T. (2014). Challenges andsolutionsformarketinginadigitalera.EuropeanManagementJournal,32(1),1–12.

Luo, X., & Bhattacharya, C. B. (2009). The debate over doing good: Corporate socialperformance, strategic marketing levers, and firm-idiosyncratic risk. Journal ofMarketing,73(6),198–213.

McGovern,G. (2007).VirginmobileUSA:Pricing for the very first time (case 9-504-208).Boston:HarvardBusinessSchoolPress.

Meer,D.(2013).TheABCsofanalytics.StrategyBusiness,70,6–8.

Neslin,S.A.,Grewal,D.,Leghorn,R.,Shankar,V.,Teeling,M.L.,Thomas,J.S.,&Verhoef,P. C. (2006). Challenges and opportunities in multichannel customer management.JournalofServiceResearch,9(2),95–112.

Onrust,M.,Verhoef,P.C.,VanDoorn,J.,&Bügel,M.S.(2014).Whendoinggoodleadstoincreased customer loyalty: Why weak firms can benefit from CSR. Working Paper,UniversityofGroningen.

Porter,M. E., & Kramer,M. R. (2011). Creating shared value.Harvard Business Review,89(1/2),62–77.

Reichheld, F. F. (1996). The loyalty effect: The hidden force behind growth, profits, andlastingvalue.USA:HarvardBusinessSchoolPress.

Reimer,K.,Rutz,O.J.,&Pauwels,K.(2014).Howonlineconsumersegmentsdifferinlong-termmarketingeffectiveness.JournalofInteractiveMarketing,28(4),271–284.

Reinartz, W. (2011). Presentation on customer management on Dutch customerperformanceawards.

Rigby,D.K.,&Ledingham,D. (2004).CRMdoneright.HarvardBusinessReview,82(11),118–129.

Rust,R.T.,Lemon,K.N.,&Zeithaml,V.A.(2004).Returnonmarketing:Usingcustomerequitytofocusmarketingstrategy.JournalofMarketing,68(1),109–127.

Rust,R.T.,Zeithaml,V.A.,&Lemon,K.N.(2000).Drivingcustomerequity:Howcustomerlifetimevalueisreshapingcorporatestrategy.NewYork:TheFreePress.

Srivastava,R.K.,Tasadduq,A.S.,&Fahey,L.(1998).Market-basedassetsandshareholdervalue:Aframeworkforanalysis.JournalofMarketing,62(1),2–18.

Taylor,L.,Cowls,J.,Schroeder,R.,&Meyer,E.T. (2014).Bigdataandpositivechangeinthedevelopingworld.Policy&Internet,6(4),418–444.

Thaler, R. H., & Tucker, W. (2013). Smarter information, smarter consumers. HarvardBusinessReview,91(1),45–54.

Van Bruggen, G. H., & Wierenga, B. (2010). Marketing decision making and decisionsupport: Challenges and perspectives for successful marketing management supportsystems.FoundationsandTrends®inMarketing,4(4),209–332.

VanDoorn, J., &Hoekstra, J. C. (2013). Customization of online advertising: the role ofintrusiveness.MarketingLetters,24,339–351.

Venkatesan, R., & Kumar, V. (2004). A customer lifetime value framework for customerselectionandresourceallocationstrategy.JournalofMarketing,68(4),106–215.

Verhoef, P. C. (2012). Klanten centraal in de bankensector. White paper, MonitoringCommissieCodeBanken,theNetherlands.

Verhoef,P.C.,&Langerak,F. (2002).Elevenmisconceptionsabout customer relationshipmanagement.BusinessStrategyReview,13(4),70–76.

Verhoef, P. C., & Leeflang, P.S.H. (2009). Understanding the marketing department’sinfluencewithinthefirm.JournalofMarketing,73(2),14–37.

Verhoef,P.C.,&Lemon,K.N.(2013).Successfulcustomervaluemanagement:Keylessonsandemergingtrends.EuropeanManagementJournal,13(1),1–15.

Wiesel, T.,Alleman, T., Bouma, J. T., Bügel,M. S., deHaan, E.,Hoving-Wesselius, T.,&Teunter, L. (2011). Customer performance impact: interessante relaties tussen DCPI,

NPS en omzet. Customer Insights Center Rapport CIC-2011-02, University ofGroningen.

2.1Value-to-customermetrics

Introduction

Value-to-customer (V2C) metrics focus on the delivered value to customers. Sometimesthesemetricsalsoreferto“shareofheart”metricsor“shareofmind”metrics.Inessence,thesemetrics indeedfocusonwhata firmachieves inacustomer’smindandwhether itresults in positive cognitive and affective responses. Thesemetrics in themselves do notreflect any value beyond what customers know and feel. However, they can indeed belinkedtovalue-to-firm(V2F)metricsandextensiveresearchhasshownsubstantialeffectsofdifferentV2CmetricsonV2Fmetrics.

In this chapterwewill discuss themainV2Cmetrics, distinguishing betweenmarket,brand, and customer metrics. Within each of these metrics, we will discuss standardmetricsandnewbigdatametrics.NexttoV2Cmetrics,wewillalsopaysomeattentiontovalue-to-society(V2S)metrics,suchascorporatesocialresponsibility.

Marketmetrics

V2Cmarket metrics aremainly relevant in the early phases of a product’s lifecycle, asdifferentfirmsaimtocommunicatethevalueandrelevanceofnewlyintroducedproductsand services. The important framework is the adoption model as proposed by Rogers(1995): he suggests that new products can be evaluated based on several dimensions:relative advantage, complexity, compatibility, observability, and trialability. In a broadersense,metricscouldfocusontheknowledgeofproducts(productawareness)andbeliefs,and on the value offered by products (product attractiveness and product uniqueness).These metrics are typically measured with surveys among potential customers usingextensive scales. In Table 2.1.1 we show an example of how these constructs are beingmeasuredfortheadoptionofanonlinegrocerychannel.Thevalidityofthesedimensionshasbeenshownfrequently,and indeedcustomerperceptionsof theseadvantagespredictusageintentionsofnewproductinnovations(e.g.,Arts,Frambach,&Bijmolt,2011;Verhoef&Langerak,2001).

Another frequentlyusedmodel in this respect is the so-called “technology acceptancemodel”(TAM).Thismodelbuildsonthetheoryofreasonedactionandsuggeststhatthereare twomainattitudes tobeconsideredfornewtechnologies:easeofuseandusefulness(e.g. Davis 1989; Davis, Bagozzi, & Warshaw, 1989). This model has also been testedfrequently formainly IT-based innovationsand itsvalidityhasbeenshown (King&He,2006).InanonlinecontexttheTAMmodelhasbeenextendedbyincludingtheeffectsoftrustandperceivedrisk(Venkatesh&Bala,2008).

Table2.1.1ExampleofitemsusedtomeasureRogers’adoptiondrivers

PerceivedrelativeadvantageElectronicshoppingislessexcitingUsingelectronicshoppingsavesmuchtimeUsingelectronicshoppingmakesmelessdependentofopeninghoursPerceivedcompatibilityElectronicshoppingsuitsmypersonElectronicshoppingrequiresfewadaptationsinmypersonallifeElectronicshoppingyieldsfewproblemsformePerceivedcomplexityElectronicshoppingiscomplex,becauseIcannotfeelandseetheproductsWithelectronicshoppingitishardtofindtheneededproductsWithelectronicshoppingitisdifficulttoorderproductsWithelectronicshoppingitisproblematictocompareproductsElectronicshoppingiscomplexIntentiontoadoptelectronicgroceryshoppingPleaseindicateontheresponsescalefrom0to10towhichextentyouintendtouseelectronicshoppingtoobtainyourgroceriesinthenearfuture

Source:AdaptedfromVerhoef&Langerak(2001)

Newbigdatamarketmetrics

Big data developments, and specifically online conversation on products and productsusage,may provide firmswith a deeper understanding of how customers view and useproducts.Of specific usehere could be statistics on theuse of different search termsonsearchenginessuchasGoogleandYahoo.Thesesearchtermsmayshowinitialinterestinproductsandbrands.

A tool used for this is Google trends. In Figure 2.1.1 we show the search results for“tablet” as a product over time. As one can observe, the number of search requests fortablets has increased over time, and one can also observe some peaks. New productintroductionsof,forexample,theiPadcouldcausethesepeaks.

Similarfigurescanbederivedfrommultiple,alsomoregeneric,searchterms.InFigure2.1.2we showsearch results forbigdataandmarket research.Hereyouclearly see thatinterestinmarketresearchisdeclining,whereasthereisastrongandincreasinginterestinbig data. Importantly, Google trends can also show the interest across, for example,differentgeographicalmarketsandspecificcitiesacrosstheglobe.

Brandmetrics

Figure2.1.1Searchresultson“tablet”worldwide

Source:AdaptedfromGoogletrends(2015)

V2Cbrandmetricsarefrequentlycollectedonacontinuousbasis.Formanyfirmsitisveryimportant tocontinuouslymeasure indicatorsof theirbrandperformanceand, related tothat, to track the outcomes of advertising campaigns. Many research and advertisingagencies around the globe, such as Young & Rubicam, have developed standard brandperformancemeasurements.Importantly,brandmetricsarecollectedamongallcustomersinthemarketplace,asfirmsaimtomeasurethepositionofbrandsrelativetocompetingbrands.Thiscontrastswithcustomermetrics,whichtypicallyaremeasuredamongexistingcustomersofafirmorbrand.

Figure2.1.2Searchinterestin“bigdata”and“marketresearch”

Source:AdaptedfromGoogletrends(2015)

Traditionalbrandperformancemeasurescanbestructuredaroundthesalesfunnelfrombeingawaretofinalpurchaseofthebrandandsubsequentresultingloyalty.Brandmetrics

can also be classified based on their focus. Broadly, one could distinguish betweencognitive brand metrics focusing on customers’ knowledge of a brand and affectivemeasures focusing on customers’ feelings and emotions towards a brand (Hanssens,Pauwels,Srinivasan,Vanhuele,&Yildirim,2014).

Atypicalcognitivebrandmetricisbrandawareness,whichmeasureswhethercustomersknowthebrand.Brandawarenesscanbeunaidedorspontaneous,reflectingtop-of-mindawarenessandaidedbrandawareness(seeFigure2.1.3foratrend-lineonthesetwobrandmetrics). Especially for strong brands, such as Coca Cola and Apple, brand awarenessmetricsarecloseto100%anddonotvarymuchovertime.Theaddedvalueofthisbrandawarenessmetricforthesebrandscouldberelativelysmall.Forunknownbrands,trackingbrandawarenesscanprovideveryusefulinformation.Beyondbrandawarenessfirmsalsofrequently measure “advertising awareness,” which focuses on whether consumers areawareof(usually)arecentbrand’sadvertisingcampaign.

At a deeper level, customers gain more knowledge about the brand. This may bereflectedinspecificbrandassociations,suchaswhetheritisaninnovativebrandorahigh-qualitybrand.Theseassociationsarelikelytovarymorebetweenbrandsandovertime,asbrands have developed specific positioning strategies. In contrast with brand awarenessmetrics, thesemetrics are likely to changemore over time (e.g.Mizik& Jacobson, 2009;Hunneman,Verhoef,&Sloot,2015).

Brandmetrics that are closer to the final purchase concern brand consideration, andbrandpreferenceorbrandliking.Brandconsiderationmetricsfocusonwhetherbrandsarein the set of brands that customers consider buying.Brandconsideration is traditionallyconsideredasanecessaryconditionforbrandstobepurchased.However,brandscaneasilyenteraconsiderationset,whentheyare,forexample,inpromotionorthereiseffectivein-storecommunication(e.g.Baxendale,Macdonald,&Wilson,2015;VanNieropetal.2010).

Figure2.1.3Exampleoftrackingaidedandspontaneousawarenessthroughtime

Brand preferencemeasureswhether customers prefer a specific brand over competingbrands. High brand preference levels for brands should typically lead to higher marketsharesforbrands.OnecouldarguethatabrandpreferencemeasureissoclosetobehaviorthatitmightactuallybeaV2Fmetric.InFigure2.1.4,weprovideanexampleofthebrandpreference for multiple smartphone brands and how these preference measures varybetween different market segments. As one can observe, the Apple iPhone is the mostpreferredbrandin2011,butthatthispreferencevariesbetweensegments.

Analternativemeasurebeingusedisbrandliking,whichisalsoametricfocusingontheaffectiveattractionofabrand. It is typicallymeasuredwithaquestion,wherecustomershavetostatetheirlikingofabrand,forexamplebyusingascalesuchas1=“notlikethebrand,”to7=“likeenormously.”(Hanssensetal.,2014).

Brand-assetvaluator®

One of themost influential V2C brandmeasurement systems is the one developed andusedbyYoung&Rubicam.TheydevelopedtheBrand-AssetValuator®(BAV).Thisbrand-assetvaluatorfocusesonmultipledimensionsofthebrand(seeFigure2.1.5).

Figure2.1.4Exampleofbrandpreferenceofsmartphoneusers,de-averagedtogenderandage

Source:AdaptedfromOfcomomnibusresearch,March2011

Figure2.1.5Brand-AssetValuator®model

Source:adaptedfromYoung&Rubicam1

The BAV distinguishes between brand strength and brand stature. Each of these

measurescanthenbesubdividedintounderlyingmetrics:differentiationandrelevanceforbrand strengthandesteemandknowledge forbrand stature. In someupdatedmodelsofBAV, brand energy is also measured (Mizik & Jacobson, 2009). The measures anddefinitions for each of these dimensions, including brand energy, are provided in Table2.1.2.

Dobrandmetricsmatter?

Collectingthebrandmetricsdiscussedaboveprovidesfirmswithearlyinformationonthefuturehealthoftheirbrands.Forexample,changesinbrandconsiderationmaysignalthatsomewhereinthenearfuturebrandmarketshare

Table2.1.2DefinitionsofBAV®components

BAVpillarUnderlyingperceptualmetrics

Surveyscale

BAVdata Meaningandroleofthepillar

Differentiation1.Unique2.Distinctive

Yes/noYes/no

%responding“yes”%responding“yes”

Perceiveddistinctivenessofthebrand.Definesthebrandandreflectsitsabilitytostandoutfromcompetition.Isthe“engineofthebrandtrain;…iftheenginestops,sowillthetrain.”

Relevance1.Relevanttome

1-7scale

Averagescore

Personalrelevanceandappropriatenessandperceivedimportanceofthebrand.Drivesmarketpenetrationandisasourceofbrand’sstayingpower.

Esteem

1.Personalregard2.Leader3.Highquality4.Reliable

1-7scaleYes/noYes/noYes/no

Averagescore%responding“yes”%responding“yes”%responding“yes”

Levelofregardconsumersholdforthebrandandvalenceofconsumerattitude.Reflectshowwellthebrandlivesuptoitspromises.

1.

Awarenessandunderstandingofthebrandidentity.Captures

Knowledge Familiaritywiththebrand

1-7scale

Averagescore

consumerintimacywiththebrand.Resultsfrombrand-related(marketing)communicationsandpersonalexperienceswiththebrand.

Energy(newpillar)

1.Innovative2.Dynamic

Yes/noYes/no

%responding“yes”%responding“yes”

Brand’sabilitytomeetconsumers’needsinthefutureandtoadaptandrespondtochangingtastesandneeds.Indicatedfutureorientationandcapabilitiesofthebrand.

Source:AdaptedfromMizik&Jacobson(2009:16)

mightdecline.Asnoted,italsoprovidesinformationonthecompetitivepositioningofthebrand compared with other brands. This information may help firms to redefinepositioning strategies and the communicated USP. For example, if the price image ofsupermarketsischanging,supermarketsmaywanttochangetheirretailmixinsuchawaythat, for example, thisprice image is improved (e.g.Hunnemanet al., 2015;VanHeerde,Gijsbrechts,&Pauwels,2008).

Research has also considered the impact of brandmetrics on V2Fmetrics.Mizik andJacobson (2009) studied the impact of the BAV®measure on shareholder valuemetrics.Theyessentiallyshowthatpositivechangesinenergyandrelevanceperceptionsincreaseshareholder value metrics. No effects on the other three dimensions are found. Theseresultssuggestthatonlyspecificbrandmetricscanbeconsideredas indicatorsforfuturebusinesssuccess.

Brandmetricshavenotonlybeenlinkedtoshareholdervaluemetrics,butalsotoV2Fmetrics,suchasbrandsales.Srinivasan,VanhueleandPauwels(2010)labelthesemetricsas“mind-setmetrics”andassessthejointimpactofpastsales,marketingmixvariables,andmind-set metrics on brands’ sales performance. Their studies consistently show thatmarketing mix variables impact these metrics and that the studied brand metrics alsoimpact sales performance. However, there are substantial differences between brands(Hanssensetal.,2014).

Hanssensetal.(2014)suggestgoodV2Cbrandmetricsshoulddowellonthreecriteria.Theyshould:

1. havepotentialforgrowth;2. havesomestickiness;and3. beresponsivetomarketingefforts.

Thefirstcriterionreferstowhetherthemetriccanindeedchangeandgrowovertime.Thisfrequentlydoesnotholdforbrandawarenessmetrics,asanaturalceilingisoftenreached.

Thestickinessofthemetricfocusesonthefactthatthemetricdoesnotchangetoomuchovertime.Thereshouldbesomestayingpower,whichmayresultfrominertiaorlock-in.Theresponsivenesstomarketingeffortsreferstomarketing’sabilityto“movetheneedle”ontheV2Cmetric.IfaV2Cmetricdoesnotrespondtochangesinthemarketingmix,itisprobablynotaveryeffectivemetric.

Whataboutbrandequity?

Acommontermusedinthebrandingworldis“brandequity.”Brandequity(BE)canhavedifferent meanings. Depending on the measurement, it is sometimes considered as justanotherV2CmetricwhereasothersconsideritmoreaV2Fmetric.BrandingexpertKevinLaneKeller considers customer-based BE as that part of customers’ behavioral responsethatcanbeattributedtothebrand.Forexample,ifcustomersarepayingapricepremiumfor a brand, this could be an indicator of BE (Ailawadi,Neslin,& Lehmann, 2003). ThedefinitionofBEindicatesthatitismainlyaV2Fmetricreflectingthe(financial)valueofthebrand.InfactfirmslikeInterbrandfocusonthisvalueofthebrand,whenmakingalistofthetop100globalbrands.

SomeresearchersmeasureBEusingattitudinalcustomermetricsthatfocusonmeasuressuchasbrandstrengthandbranduniqueness(Verhoef,Langerak,&Donkers,2007).Infact,in his BEmodel, early brand guru David Aaker includes brand associations and brandquality as indicators of BE (Aaker, 2009). Brand associations could be quality,innovativeness, but also associations, such as hip, young, etc. Examples of brands withstrongbrandassociationsincludeBMWandNike.

Newbigdatabrandmetrics

Themetricsdiscussedso fararerather traditionalandhavebeenaroundforyears.Onlyrecently we have started to understand the actual impact of these metrics on brandperformance outcomes. Being able to combine different data sources, together withcontinuouscollectionofbrandmetricsdata,allowedresearcherstodoso.

Particularlyasaresultofonlineandsocialmediadevelopments,wherecustomerssharetheiropinionsaboutbrandsandmayalsoindicatetheirlikingofbrands,newsourcesfordata on brands have been developed. This is sometimes referred to as “user generatedcontent” (UGC). Importantly, this UGC can be collected and analyzed to create brandmetrics.Weconsiderthefollowingspecificnewbigdatametrics:

DigitalbrandassociationnetworksSummarizeddigitalbrandmetricsSocialmediabrandmetrics.

Digitalbrandassociationnetworks

Whenwritingtheiropinionaboutbrands,customersmaysharedifferentopinionsinwordsaboutbrands.Forexample,theymaysharesomewhereonablogthatAppleisconsideredinnovative, whereas Samsung is considered as a real Asian brand. Similarly, ideas onbrands such asNike andAdidas can be shared. Researchers have developedmethods tocollectthesedataandtoanalyzethem,therebyconsideringthevalenceofthesewords.InFigure2.1.6anassociationnetworkforMcDonaldsisshown,basedondigitaldata.Asonecanobservefromthisnetwork,themainassociationforMcDonaldsis“yummy”;however,negative associations are service, taste and the offered volume (not enough). We willdiscussmethodstoanalyzetextdataingreaterdetailinChapter3.1.

Digitalsummaryindices

Withdigitalbrandassociationswearemainlyinterestedinactualassociations,providingmanagerswithmoreknowledgeonwhatcustomersactuallythinkaboutabrand.Researchagencieshavenowdevelopedmethodstoassesshowpositiveornegativecustomers’viewsarewhen they discuss brands online. Using text analytics, the valence ofwords is thenassessed. Based on dictionaries, such as the dictionary of affect, the negativity andpositivityofthesewordscanbeassessed.Whereasinthepastthiswasdonemanually(e.g.Verhoef,Antonides,&DeHoog2004), this isnowdoneautomaticallyusingtextanalysisprograms.Basedonthispositiveandnegativevalenceforbrands,avalencescorecanbecalculated. These valence scores are related to the sales of brands; correlations with

shareholdervaluemetricshavealsobeen reported (e.g.Tellis& Johnson, 2007;Onishi&Manchanda,2012).Thesedigitalsummaryindicesarealsoreferredtoaselectronicword-of-mouth,oreWOM(Trusov,Bucklin,&Pauwels,2009).

Figure2.1.6AssociationnetworkofMcDonald’sbasedononlinedataa

Source:AdaptedfromGensler,Völckner,Egger,Fischbach,&Schoder(2015)

NoteaGrey (white) circles respresent favorable (unfavorable) brand associations.Numbers incirclesrepresentnormalized,weighteddegreecentrality(permill)

Especially in the last decade, specialized companies have developed specific brandmetrics that can summarize customer online discussions. One example is the digitalsentiment index(DSI)asdevelopedbyOxyme/Metrixlab incooperationwithresearchersfromtheUniversityofMünster.TheDSIincorporatesthesentimentconcerningproducts,services, or brands throughout the most important platforms in a single measure.Furthermore, DSI also tracks competitors and one can simply derive direct comparisonswiththem.DSIisusedformorethanahundredbrandsinGermany,USA,UK,France,theNetherlands,Finland,andSweden.

Socialmediabrandmetrics

Not only do customers discuss brands online in social media, but also the brandsthemselves actively use socialmedia for promotion purposes. Customers can visit thesebrand pages on, for example, Facebook. Customers can react to these brand pages bypushingonthe“like”buttonandbyprovidingcomments.Thenumbersofbrandlikesand

brandcommentsareconsidered tworelevantsocialmediabrandmetrics.Thenumberofbrand likes may be an indicator of brand preference among customers. The number ofcommentsmayindicatesomebrandinvolvement.Importantly,thecontentofsocialmediamarketingcampaignsaffectsbothmetrics.Forexample,brandsgetmorelikeswhentheyincludeacontestinthecampaignandincludeavideo(DeVries,Gensler,&Leeflang,2012).Not surprisingly, more comments are received when a question is being asked in thecampaign.Overall,onecoulddoubtthevalueofthesemetricsasrealV2Cbrandmetrics.Still,thenumberoflikesandcommentscanbesubstantial,althoughitvariesconsiderablybetween brands and industries (see Figure 2.1.7). To some extent, they mainly reflectreactions to the social media presence of a specific brand, while brand knowledge andattitudesarebasedonmultipleinteractionsinmultiplechannelsandtouchpoints.Finally,wenotethatsocialmediabrandmetricsandthecontentofdiscussionsonsocialmediaofbrandsmaycontributetothesummarydigitalbrandmetrics,suchasDSI.

Customermetrics

V2Ccustomermetrics are frequently labeledas customer feedbackmetricsbymarketingresearchers. Customer feedback metrics (CFMs) have become very popular. Firms areaiming to improve thecustomerexperienceacrossmultiple touchpointsandare seekingways of measuring this experience. Targets are set on these metrics and continuousmeasurement occurs. In some cases specific feedback mechanisms are built in. Such amechanismistheso-calledcustomerfeedbackloop.Inthisloopcustomersgivefeedbacktoafirmonaspecificserviceeventandsubsequentlytheyarecalledbackwhentheyscorelowonthismetric,andareaskedaboutwhat isbehind this lowscore.Firms then try tosolve these issues, which hopefully results in a higher performance. All these kinds ofsystemsresult inseriesofcustomerfeedbackmetricsof thousandsofdifferentcustomersover time. The most popular metrics are the net promoter score (NPS), customersatisfaction, and the more recent customer effort score (CES). These metrics can bemeasuredindifferentways.InTable2.1.3theexactquestionsandoperationalizationsarediscussed:

Figure2.1.7Averagenumberoflikesandcommentsperproductcategory

Source:AdaptedfromDeVries,Gensler,&Leeflang(2012:87)

We distinguish between these metrics on two dimensions. The first dimension wasintroduced by Bolton, Lemon and Verhoef (2004) andmore recently by Zeithaml, et al.(2006),whofocusonthetimespanofmeasuresanddistinguishbetweenmorebackward-looking and more forward-looking metrics. Forward-looking CFMs focus on whatcustomersplantodointhefutureandmaysignalsomethingaboutthefutureperformanceof therelationship.TheNPS introducedbyReichheld (2003) isanexampleofa forward-lookingCFM,sinceitconsidersthewillingnesstorecommendafirminthefuture,whichmayalsosignalacustomer’s futurerelationshipwith the firm(e.g.Zeithamletal.2006).Backward-looking metrics focus on the past and current customer performance of acompanytowardscustomers.

The CES is a typical backward-looking CFM, as it measures the perceived service

performancefromapastspecificexperience(Dixon,Freeman,&Toman,2010).Itisbasedonasinglequestion(“Howmucheffortdidyoupersonallyhavetoputforthtohandleyourrequest?”),andmeasuredonafive-pointscale.Dixonetal.(2010)suggestthattheCESisabetter predictor of repurchase (intentions) and increased spending than the NPS orcustomersatisfaction.

Table2.1.3Overviewofdifferentcustomerfeedbackmetrics

CFM Measurement

1.Customersatisfaction

“Allinall,howsatisfiedorunsatisfiedareyouwith[companyX]?”(1=veryunsatisfied,7=verysatisfied).

2.Top-2-boxcustomersatisfaction

Adummyatthecustomerlevelindicatingifthecustomerhasgivenascoreofsixorsevenonthecustomersatisfactionquestion.Atthefirm(industry)levelthisistheproportionofcustomersofthatfirm(industry)thatgaveascoreof6or7.

3.Netpromoterscore(NPS)

“Howlikelyisitthatyouwouldrecommend[companyX]toafriendorcolleague?”(0=veryunlikely,10=verylikely).Respondentswhogaveascoreof0-6are“detractors,”thosewhogavea7or8are“passives,”andthosewhogavea9or10are“promoters.”SubtractingtheproportionofpromotersfromtheproportionofdetractorsprovidestheNPSatthefirmlevel(Reichheld,2003).

4.NPS-value

TheuntransformedNPSscore(onthe0-10range)providedbythecustomer.

5.Customereffortscore(CES)

“Didyoutrytocontact[companyX]withanykindofrequest?”(yes/no)Ifyes,thefollowingquestionisasked:“Howmucheffortdidyoupersonallyhavetoputforthtohandleyourrequest?”(1=veryloweffort,5=veryhigheffort).

Source:AdaptedfromDeHaan,Verhoef,&Wiesel(2015)

Finally,customersatisfactionfocusesmoreonanoverallevaluationof the interactionsbetween the customer and the firm over time, and tends to have amore present focus(Verhoef,2003),althoughitmayalsobebasedonpastexperiences.

TheseconddimensionweuseisabouthowthemeasurementscaleoftheCFMisused.Thereareadvocatesofnotlookingatthemeanvalueofthescale,butattheproportionofpeoplerespondingverypositivelyand/orverynegatively.Anexampleofthisis“top-2-boxcustomer satisfaction,” which measures the proportion of customers filling in the twohighest-scoringpointsfortheiroverall2customersatisfaction(Morgan&Rego,2006).ThecalculationunderlyingtheofficialNPSalsodistinguishesbetweenverypositive,moderate,andnegativeresponses(Reichheld,2003).Transformationscantheoreticallybedefended,asithasbeenshownthatcustomersmainlyfocusonextremeexperiencesandthereforetheeffectsofCFMscanberathernon-linear(e.g.VanDoorn&Verhoef,2008;Streukens&De

Ruyter,2004).Moreover,servicemarketingexpertspromisetodelightcustomers,implyingthatcustomersshouldevaluatefirmswithextremescoresontheCFMscales(Oliver,Rust,&Varki,1997).Firmscan,however,alsochoosenot tousea transformation,and insteadmakeuseofthefullscale,forexamplethe0–10scaleoftheNPS.Ifwecombinethetwodimensions,we end upwith the three-by-two classificationmatrix as provided in Table2.1.4.

Table2.1.4Conceptualizationofcustomerfeedbackmetrics

Pre-defineddata

Pastfocus PresentfocusFuturefocus

Partofthescaleused

Fullscale Customereffortscore(CES)

Customersatisfaction NPSValue

Focusonextremes

Top-2-boxcustomersatisfaction

OfficialNPS

Source:AdaptedfromDeHaan,Verhoef,&Wiesel(2015)

Isthereasilvermetric?

Proponentsofdifferentmetricspropose that theirmetricshave thebestperformance forpredicting future growth and customer retention. For example Reichheld (2003) stronglyadvocatestheNPS,whileDixonetal.(2010)believeandshowthatCEShasaverystrongperformance in predicting customer loyalty, going beyond the performance of othercompeting metrics. Not surprisingly, academics questioned these claims and haveinvestigatedtheactualqualityofthesemetrics.Typicallytheycomparedtheperformanceof different metrics in predicting future business growth and customer loyalty. Initialstudies were not so positive about the performance of the NPS and tended to prefercustomer satisfaction (e.g. Keiningham, Cooil, Aksoy, Andreassen, & Weiner, 2007).However, more recent studies actually show smaller differences between customersatisfactionandNPS(e.g.VanDoorn,Leeflang,&Tijs,2013).Inarecentstudyweanalyzedtheperformanceofthemetricsforcustomerretention(DeHaanetal.,2015).Hereweagainfind no clear winner. NPS and customer satisfaction score equally well. The CES is,however, doing very poorly. We also observe that there might be some benefits incombiningmetrics,suggestingthatatleastfirmsshouldmonitormultiplemetrics(e.g.NPSandsatisfaction).Thiswouldimplymakinguseofdashboardsinvolvingmultiplemetrics.Based on existing knowledge and practical insights we have the followingrecommendationsforfirms:

Relyonoverallmetricswithapresentorfuturefocus.Itisveryvaluabletoanalyzethedevelopmentofthetopscoresofmetrics(e.g.top-2boxes).Do not immediately adopt new metrics promising superior performance, butcarefullyconsidertheactualperformance.Measuremultiplemetricsandreportinadashboard.

Othertheoreticalrelationshipmetrics

Especially within the CRM literature, other metrics or customer attitudes have beendiscussed and have gained attention. These metrics mainly focus on the quality of thecustomerrelationship.Customersatisfactioncanbean indicatorof thisquality,althoughthisattitudeusuallyfocusesmoreonthecognitivesideoftherelationship,beingthemereevaluation of the delivered products and services (Bolton et al., 2004). Themost specificmetricthathasbeenproposediscommitment,definedastheenduringdesiretocontinueavalued relationship (Moorman, Zaltman, & Desphande, 1992). In general this attitudereflects a more emotional evaluation and also considers the future development of therelationship.Researchershavealsoconsideredseveralformsofcommitment(e.g.Verhoef,Franses,&Hoekstra,2002):

Affectivecommitment: psychological attachmentof a customer to the firm,basedon feelings of identification, loyalty, and affiliation (Gundlach, Ravi, &Mentzer,1995).Calculative commitment: the extent to which customers perceive the need tomaintainarelationshipgivenanticipatedterminationorswitchingcosts(Geyskens,Steenkamp,Scheer,&Kumar,1996).Normative commitment: a customer’s obligation-based attachment to anorganization,markedbyfeelingsofguiltoruneaseinleavingtheorganization(e.g.,Melancon,Noble,&Noble,2011).

Affectivecommitmenthasbeenshowntobeapredictorofretention,shareofwalletandword-of-mouth (Verhoef, 2003). The two other forms of commitment arise frommerelynegativemotivationstostayinarelationshipandcanevenbedetrimentalforrelationshipdevelopment(Verhoefetal.,2002).

Interestingly,basedonthetheoryoflovedevelopedbythepsychologistR.J.Sternberg(1986),Bügel,VerhoefandBuunk(2011)propose“customerintimacy”asanovelcustomerattitude.Intheiranalysisthismeasureinvolvesbothpassionabouttherelationshipandtheperceived intimacy between the customer and the firm. Their results suggest a strongcorrelationbetweencommitmentandcustomerintimacy.However,customersfeelstrongercommitmentthanintimacy.Remarkably,theyalsoshowthatthelevelofintimacydevelops

during the customer lifecycle from relatively high in the very early phases of therelationship,whileitdropsinlaterphasesbeforerisingagain(seeFigure2.1.8).Overall,weobservenostronguseofthesemetricswithinpractice,despitethefactthattheyhavebeenverywellreceivedwithinacademicresearch.Itseemsthatthesemeasuresareconsideredhighly theoretical, without a strong adoption within marketing practice (e.g., Roberts,Kayandé,&Stremersch,2014).

Another customer metric that gained attention is customer trust. Trust is defined ascustomers’confidenceinthequalityandreliabilityoftheservicesprovided(Garbarino&Johnson, 1999). Researchers have also proposed that one should distinguish betweenreliabilityandbenevolence.Reliabilityfocusesonthefactthatafirmactsonitspromises,whilebenevolenceconsidersthefactthatafirmnotonlycaresaboutafirm’sinterests,butalso the customer’s interest, and acts on that (Geyskens et al., 1996). Trust has gainedrenewedattentionasaresultofthemultiplecrisesfirmsandcustomershavebeen(andstillare)confrontedwith.Inparticularthefinancialcrisishasstirredupdistrustinbanksandthebankingsector(Verhoef,2012).

Figure2.1.8Developmentofintimacyandcommitmentovertime

Source:AdaptedfromBügel,Verhoef,&Buunk(2011:253)

Afinaltheoreticalmetricis“paymentequity”or“pricefairness,”definedascustomers’perceived fairness of the price paid for their consumed services of products (Bolton &Lemon, 1999). In contrast with other metrics, this metric strongly focuses on a singleinstrument: price. It has been shown that it can predict service usage. However, otherstudies have shown its predictive power for retention and the purchase of additionalservicestobeabsent(Verhoef,2003;Verhoefetal.,2002).

Customerequitydrivers3

WithincustomermanagementthecustomerequitymodelasdevelopedbyRust,Zeithaml,and Lemon (2000) has been influential. They consider customer equity—the net presentvalue of all future and current customers—as a very important outcome variable incustomer-centricfirms.Conceptuallytheyconsiderthreedrivers:

value equity (VE), defined as customers’ objective assessment of the utility ofservices based on perceptions of “what is given up” for “what is received.” VEreflects the outcome of customers’ comparisons between their own expectationsandfirms’performance;brandequity(BE),whichreflectscustomers’subjectiveandintangibleassessmentofthebrandimage;andrelationship equity (RE), defined as customers’ assessments of their interactionswith the firm. This factor depends on customers’ relationships with sales andservice persons, loyalty programs, customer communities/networks, and so forth.Positive RE provides relatively more financial and social benefits to customers(Rust, Zeithaml & Lemon, 2000). This enhances feelings of reciprocity andbenevolence,whichshouldpositivelyinfluenceloyalty(Selnes&Gønhaug,2000).

These drivers can bemeasuredwith attitude-like questions inwhich customers evaluatetheperformanceoffirmsandbrandsonthesedimensions.Thereissufficientevidencetoindicatethatthesedriverscontributetocustomerloyalty.However,theimportanceofeachof these driversmay differ between industries, firms, and customers (e.g. Ou,De Vries,Wiesel, & Verhoef, 2014). For example, BE and VE aremore relevant for firms offeringinnovativeservices.

Newbigdatacustomermetrics

Big data customer V2C metrics mainly arise from online interactions, but also frominternaldatasources,whicharefrequentlyneglected.

Internaldatasources

Within CRM databases the focus has been strongly on collecting transaction data.However,firmshavemanyinteractionswithcustomersthatcanbeusedasindicatorsforV2C.Ingeneralweadvocatepayingmoreattentiontothisinformation.Forexample,firmscan collect complaints of customers. A complaint is certainly a V2C indicator and canprobably predict switching behavior. We note that, according to the service recoveryparadox,firmsactingoncomplaintscanincreasecustomerloyalty(VanDoorn&Verhoef,2008). Also personal contacts with customers can be indicators of V2C. More deeplypersonal contacts (e.g. email-conversations, call centercontacts) canbeanalyzedand thepositivityornegativityofthesecontactscanbeassessed(Verhoef,Antonides,&DeHoog,2004).

Other internal data sources can provide information on the actual delivered value tocustomers and specifically use internal operations data. These data can be customer-specific ormeasured at amore aggregate level.An example of a customer specific datamightinvolvetheresolutiontimeofaserviceproblemforaspecificcustomer.Boltonetal.(2008) show that operational data can indeedpredict the probability of contract renewalandthelevelofserviceupgrading.Thedeliveredserviceperformancecanalsobemeasuredatamoreaggregate level.Forexample, for railwayfirms thepercentageof trainswithaspecific delay or the percentage of unsuccessful connections between trains can be aninternal measure for the actual delivered value (Gijsenberg et al., 2015). One importantissue is that these internalmeasuresdonot reflect theactualperceptionof thedeliveredvalue.Perceptionsfrequentlyarisefromacomparisonoftheexpectedservicelevelandthedelivered service level.Expectationsare,however, frequentlybasedon thepastand thushaving time series data on the delivered performance may create a more accurateunderstanding(Gijsenbergetal.,2015).

Onlinesources

The most important online customer metric is customer reviews. Reviews of firms andspecificproductsareplacedonlineandfrequentlyinvolverelativelyindependentwebsites,suchasTripadvisororZoover.Also,onlineretailersallowconsumerstoprovidereviewsofsoldproducts.Anonlinesurveyof2,005Americanshoppersshowedthat65%ofpotential

shoppers selected a brand that was only in their consideration set as a result of onlinereviews (Weber Shandwick, 2012). This suggests that reviews can be very powerful andinfluential in steering customer choices online. Indeed, the effect of reviews on sales issignificantandsubstantialwithameanelasticityof0.35 increase forreviewvolumeand0.65 for review valence (Floyd, Freling, Alhoqail, Cho, & Freling, 2014). Here “reviewvolume”isameasureforthenumberofreviewsand“reviewsvalence”isameasureofthepositivity of the review. However, there is considerable variation in the effectiveness ofreviews. Reviews have stronger predictive effects for products with a high customerinvolvement, while third party and critic reviews have a stronger effect than normalcustomerreviews.

Note that in an online context the differences between brand metrics and customermetrics become blurred. Customer reviews can be used as input for the creation ofsummary brand metrics. Furthermore, customer reviews also are considered as eWOM.Despitethisitisratherclearthatreviewsarebecomingveryimportant.Theyclearlyaffectsalesascustomersincludereadingthesereviewsintheirpurchasedecision.If,forexample,ahotelhaspoorreviews,fewercustomerswillbelikelytobookthathotel.

V2Smetrics4

TheV2Cconceptcanalsobeextendedtofocusonthesocietalvalueoffirms.Manyfirmsaredevelopingstrategiesandinitiativesto,forexample,improvetheirsustainability.Suchinitiatives are significant elements of the corporate strategy of many (multi) nationalcorporations(Beardetal.,2011;Porter&Kramer,2011),whoseCEOs,suchasPaulPolmanofUnilever,assertthatbusinessescanbeapositiveforceforgoodintheworldandthatthisapproachisintheinterestsofallfirms’stakeholders.

Firms also collect metrics to measure their performance on these societal strategies.These metrics are typically not the responsibility of marketing, but mainly areadministered by staff departments responsible for sustainability and/or corporatereputation.Withinthemanagementandmarketingliterature,thereisonemetricreceivingconsiderableattention:corporatesocialresponsibility(CSR).

Corporatesocialresponsibility

CSR is a firm’s commitment to ensure societal and stakeholder well-being throughdiscretionarybusinesspracticesandcontributionsofcorporateresources(Du,Bhattacharya&Sen,2010;Kotler&Lee,2005;Luo&Bhattacharya,2006).Asabroadconcept,CSRcanincludebusiness practices asdiverse as cashdonations to charity, equitable treatmentofworkers, and an environmentally friendly production policy. Yet, although CSR goesbeyondtheinterestsofthefirmtobenefitsociety(McWilliams&Siegel,2001),manyfirmsadditionallystriveto“dobetterbydoinggood”andgaincompetitiveadvantagesthroughCSR (Prout, 2006). Frequently CSR is seen as a kind of perception measure similar tosatisfaction, and involves seeking opinions on statements such as “this companyemphasizes the importance of its social responsibilities” and “this company provides anevident social contribution” (Du, Bhattacharya & Sen, 2007). There is evidence thatcompanies performingwell onCSRhave a higher performance.However, the effects oncustomerbehavior(i.e.retention)arelessstraightforward.InfactCSRisbelievedtohavean indirect effect on loyalty through improving customerattitudes (e.g.Onrust,Verhoef,Van Doorn, & Bügel, 2014). From a big data perspective marketing analytics should beawarethatthesedataarefrequentlycollectedwithinfirms.However,asnotedearlier,thesedataaretypicallynotownedbymarketingbutbycorporateleveldepartments.

Corporatereputation

Corporatereputationisarathergeneralmetricthatmeasuresthereputationoffirmsatthecorporatelevel.Itisahigh-levelmetricthatinvolvesmultiplestakeholders.Fombrunand

VanRiel (1997)definecorporate reputationasacollective representationofa firm’spastactionsandresultsthatdescribesthefirm’sabilitytodelivervaluedoutcomestomultiplestakeholders. It gauges a firm’s relative reputation both internally with employees andexternally with its stakeholders, in both its competitive and institutional environments.Severalmeasurementsystemshavebeendeveloped tomeasurecorporatereputation.TheReputation Institute has developed RepTrack®,whichmeasures corporate reputations offirms across the globe. These measures are being used by leading firms. RepTrack®involves examining the relationship between the emotional connection, or “Pulse,” withanygiven stakeholder, alongsideperceptionsof sevenunderlying rational connectionsor“dimensions” identified as: (1) products/services; (2) innovation; (3) workplace; (4)citizenship;(5)governance;(6) leadership;and(7)performance.Insideeachdimensionliespecific attributes that can be customized for clients to allow for programandmessage-readyanalysis.5Themaindifferencebetweenreputationmetricsandbrandmetricsisthatreputationmetricsaremeasuredat the firmlevelandnotat thebrand level. Inaddition,brandmetricsusuallyfocusonlyoncustomersasastakeholder,whereasreputationmetricsinvolvesmultiplestakeholders.

ShouldfirmscollectallV2Cmetrics?

A reader of this in-depth chapter might wonder whether firms should collect all thementionedmetrics.Our answer to this is adefiniteNo! Firms should focusona limitednumber of metrics and include them in a marketing dashboard. Managers should thenstrive to influence thesemetrics withmarketing strategies.We specifically observe thatmanyfirmsarenotfollowingthispath.TheycollectaplethoraofV2Cmetricsindifferentlayersintheorganization.Forexample,manyfirmscollectsatisfactiondata,dataonNPS,brand-metrics,digitalsentimentmetrics,andcorporatereputationmetrics.SatisfactionandNPS data are mainly under the ownership of the marketing research department.Marketing management and communication are mainly interested in brand metrics,whereas corporate strategy owns the corporate reputation metrics. The use of multiplemeasuresresultinstrongdiscussionsonwhichmetricstouse,insteadofhowtheycouldbeinfluencedandhowtheirusecouldgivevaluetothefirm.Whenthesemetricsarebeinganalyzed it frequently becomes clear that there are actually strong correlations betweenmany of them. An in-depth comparison of each of these metrics using specific criteriadevelopedbyAilawadietal.(2003)onwhatgoodmetricsare(seeTable2.1.5)canbeusefultoselectareducedsetofV2Cmetrics.

Conclusions

Table2.1.5Criteriaforgoodmetrics

1Theory-based2Complete3Diagnostic4Futurepotential5Objective6Basedonexistingdata7Asinglenumber8Intuitiveandtrustworthyfortopmanagement9Robustandreliable10Validatedwithotheroutcomemeasures

Source:AdaptedfromAilawadi,Neslin,&Lehmann(2003)

InthischapterwehaveprovidedanoverviewofthemainV2Cmetrics.Mostattentionhasbeen given to the metrics at the brand and customer level. These metrics have beendeveloped strongly in science and practice. Importantly, big data developments haveenriched the set ofmetrics.Themaindevelopmenthere is the fact that customers sharetheirthoughtsandfeelingsofbrandandservicesonline.Thisresultsinmanynewmetrics,whichcanbeusefulformanagers.However,someofthesemetricsarealsoratherspecific.Further, firms should consider internal data sources that can give information on theprovisionofV2C.Thesemetricscanbepowerful indicatorsofV2C.Finally,wediscussedV2Smetrics,whicharefrequentlycollectedatthecorporatelevel.

Notes

1Seehttp://young-rubicam.de/tools-wissen/tools/brandasset-valuator/?lang=en(accessedSeptember15,2015).

2ThispartofthetextisderivedfromDeHaanetal.(2015).

3ThissectionisbasedonOu,DeVries,WieselandVerhoef(2014).

4ThissectionispartiallybasedonOnrustetal.(2014).

5www.reputationinstitute.com/about-reputation-institute/the-reptrak-framework.

References

Aaker, D. A. (2009).Managing brand equity. Simon and Schuster. New York: The FreePress.

Ailawadi,K.L.,Neslin,S.A.,&Lehmann,D.R.(2003).Revenuepremiumasanoutcomemeasureofbrandequity.JournalofMarketing,67(4),1–17.

Arts, J., Frambach, R. T., & Bijmolt, T. H. A. (2011). Generalizations on consumerinnovation adoption: A meta-analysis on drivers of intention and behavior.InternationalJournalofResearchinMarketing,28(2),134–144.

Baxendale, S., Macdonald, E., Wilson, H. N. (2015). Impact of different touchpoints onbrandconsideration.JournalofRetailing,91(2),235–253.

Beard,A.,Hornik,R.,Wang,H.,Ennes,M.,Rush,E.,&Presnal,S. (2011). It’shard tobegood.HarvardBusinessReview,89(11),88–96.

Bolton,R.N.,&Lemon,K.N. (1999).Adynamicmodelof customers’usageof services:Usage as an antecedent and consequence of satisfaction. Journal of MarketingResearch,36(2),171–186.

Bolton, R. N., Lemon, K. N., & Verhoef, P. C. (2004). The theoretical underpinnings ofcustomer asset management: A framework and propositions for future research.JournaloftheAcademyofMarketingScience,32(3),271–292.

Bolton, R. N., Lemon, K. N., & Verhoef, P. C. (2008). Expanding business-to-businesscustomer relationships: Modeling the customer’s upgrade decision. Journal ofMarketing,72(1),46–64.

Bügel,M.S.,Verhoef,P.C.,&Buunk,A.P.(2011).Customerintimacyandcommitmenttorelationships with firms in five different sectors: Preliminary evidence. Journal ofRetailing&ConsumerServices,18(4),247–258.

Davis, F. D. (1989). Perceived usefulness, perceived ease of use, and user acceptance ofinformationtechnology.MISQuarterly,13(3),319–340.

Davis, F. D., Bagozzi, R. P., & Warshaw, P. R. (1989). User acceptance of computertechnology:Acomparisonoftwotheoreticalmodels.ManagementScience,35(8),982–1003.

DeHaan,E.,Verhoef,P.C.,&Wiesel,T.(2015).Thepredictiveabilityofdifferentcustomerfeedbackmetrics for retention. International Journal ofResearch inMarketing,32(2),195–206.

DeVries,L.,Gensler,S.,&Leeflang,P.S.H.(2012).Popularityofbrandpostsonbrandfanpages:Aninvestigationoftheeffectsofsocialmediamarketing.JournalofInteractiveMarketing,26(2),83–91.

Dixon, M., Freeman, K., & Toman, N. (2010). Stop trying to delight your customers.HarvardBusinessReview,88(7/8),116–122.

Du, S., Bhattacharya, C. B.,& Sen, S. (2007). Reaping relational rewards from corporatesocial responsibility: The role of competitive positioning. International Journal ofResearchinMarketing,24(3),224–241.

Du, S., Bhattacharya, C. B., & Sen, S. (2010). Maximizing business returns to corporatesocial responsibility (CSR):The role ofCSR communication. International Journal ofManagementReviews,12(1),8–19.

Dwyer,F.R.,Schurr,P.H.,&Oh,S.(1987).Developingbuyer-sellerrelationships.JournalofMarketing,51(2),11–27.

Floyd, K., Freling, R., Alhoqail, S., Cho,H. Y., & Freling, T. (2014). How online productreviewsaffectretailsales:Ameta-analysis.JournalofRetailing,9(2),217–232.

Fombrun, C. J., & Van Riel, C.B.M. (1997). The reputational landscape. CorporateReputationReview,1(1/2),5–13.

Garbarino, E., & Johnson, M. (1999). The different roles of satisfaction, trust, andcommitmentincustomerrelationships.JournalofMarketing,63(2),70–87.

Gensler, S., Völckner, F., Egger, M., Fischbach, K., & Schoder, D. (2015). Listen to yourcustomers: Insights into brand image using online consumer-generated productreviews.InternationalJournalofElectronicCommerce,20(1),112–141.

Geyskens,I.,Steenkamp,J.-B.E.M.,Scheer,L.K.,&Kumar,N.(1996).Theeffectsoftrustandinterdependenceonrelationshipcommitment:Atransatlanticstudy.InternationalJournalofResearchinMarketing,13(4),303–317.

Gijsenberg,M.J.,VanHeerde,H.J.,&Verhoef,P.C.(2015).Lossesloomlongerthangains:Modeling the impact of service crises on customer satisfaction over time. Journal ofMarketingResearch,52(5),642–656.

Gundlach, G. T., Ravi, S. A., & Mentzer, J. T. (1995). The structure of commitment inexchange.JournalofMarketing,64(3),34–49.

Hanssens, D. M., Pauwels, K. H., Srinivasan, S., Vanhuele, M., & Yildirim, G. (2014).Consumer attitude metrics for guiding marketing mix decisions.Marketing Science,33(4),534–550.

Hunneman,A.,Verhoef,P.C.,&Sloot,L.M.(2015).Theimpactofconsumerconfidenceonstoresatisfactionandshareofwalletformation.JournalofRetailing,91(3),516–532.

Keiningham,T.L.,Cooil,B.,Aksoy,L.,Andreassen,T.W.,&Weiner,J.(2007).Thevalueofdifferent customer satisfaction and loyalty metrics in predicting customer retention,recommendation,andshare-of-wallet.ManagingServiceQuality,17(4),361–384.

King, W. R., & He, J. (2006). A meta-analysis of the technology acceptance model.Information&Management,43(6),740–755.

Kotler,P.,&Lee,N.(2005).CorporateSocialResponsibility—Doingthemostgoodforyourcompanyandyourcause.NewJersey:JohnWileyandSons.

Luo,X.,&Bhattacharya,C.B.(2006).Corporatesocialresponsibility,customersatisfaction,andmarketvalue.JournalofMarketing,70(4),1–18.

McWilliams,A.,&Siegel,D.(2001).CorporateSocialResponsibility:Atheoryofthefirmperspective.AcademyofManagementReview,26(1),117–127.

Melancon,J.P.,Noble,S.M,&Noble,C.H.(2011).Managingrewardstoenhancerelationalworth.JournaloftheAcademyofMarketingScience,39(3),341–362.

Mizik,N.,&Jacobson,R.(2009).Valuingbrandedbusinesses.JournalofMarketing,73(6),137–153.

Moorman,C., Zaltman,G.,&Deshpande, R. (1992). Relationship between providers andusers of market research: The dynamics of trust within and between organizations.JournalofMarketingResearch,29(3),314–328.

Morgan,N.A.,&Rego,L.L.(2006).Thevalueofdifferentcustomersatisfactionandloyaltymetricsinpredictingbusinessperformance.MarketingScience,25(5),426–439.

Oliver,R.L.,Rust,R.T.,&Varki,S. (1997).Customerdelight:Foundations, findings,andmanagerialinsight.JournalofRetailing,73(3),311–336.

Onishi,H.,&Manchanda,P. (2012).Marketingactivity,bloggingandsales. InternationalJournalofResearchinMarketing,2(3),221–234.

Onrust,M.,Verhoef,P.C.,VanDoorn,J.,&Bügel,M.S.(2014).Whendoinggoodleadstoincreased customer loyalty: Why weak firms can benefit from CSR. Working paper,UniversityofGroningen.

Ou,Y.-C.,DeVries,L.,Wiesel,T.,&Verhoef,P.C.(2014).Theroleofconsumerconfidenceincreatingcustomerloyalty.JournalofServiceResearch,17(3),229–354.

Porter,M. E., & Kramer,M. R. (2011). Creating shared value.Harvard Business Review,89(1/2),62–77.

Prout, J. (2006).Corporate responsibility in the global economy:Abusiness case.SocietyandBusinessReview,(2),184–191.

Reichheld,F.F.(2003).Theonenumberyouneedtogrow.HarvardBusinessReview,81(12),46–54.

Roberts,J.H.,Kayande,U.,&Stremersch,S.(2014).Fromacademicresearchtomarketingpractice: Exploring the marketing science value chain. International Journal ofResearchinMarketing,3(2),127–140.

Rogers,E.M.(1995).Thediffusionofinnovations.NewYork:TheFreePress.

Rust,R.T.,Zeithaml,V.A.,&Lemon,K.N.(2000).Drivingcustomerequity:Howcustomerlifetimevalueisreshapingcorporatestrategy.NewYork:TheFreePress.

Selnes,F.,&Gønhaug,K.(2000).Effectsofsupplierreliabilityandbenevolenceinbusinessmarketing.JournalofBusinessResearch,49(3),259–271.

Srinivasan, S., Vanhuele, M., & Pauwels, K. (2010). Mindset metrics in market responsemodels:Anintegrativeapproach.JournalofMarketingResearch,47(4),672–684.

Sternberg,R.J.(1986).Atriangulartheoryoflove.PsychologicalReview,93,119–135.

Streukens, S., & De Ruyter, K. (2004). Reconsidering nonlinearity and asymmetry incustomersatisfactionandcustomer loyaltymodels:Anempiricalstudyinthreeretailservicesettings.MarketingLetters,15(2/3),99–111.

Tellis,G.J.,&Johnson,J.(2007).Thevalueofquality.MarketingScience,26(6),758–773.

Trusov, M., Bucklin, R. E., & Pauwels, K. (2009). Effects of word-of-mouth versustraditional marketing: Findings from an Internet social networking site. Journal ofMarketing,73(5),90–102.

VanDoorn,J.,&Verhoef,P.C.(2008).Criticalincidentsandtheimpactofsatisfactiononcustomershare.JournalofMarketing,72(4),123–142.

Van Doorn, J., Leeflang, P.S.H., & Tijs, M. (2013). Satisfaction as a predictor of futureperformance:Areplication.InternationalJournalofResearchinMarketing,30(3),314–318.

VanHeerde,H.J.,Gijsbrechts,E.,&Pauwels,K.(2008).Winnersandlosersofamajorpricewar.JournalofMarketingResearch,45(5),499–518.

VanNierop, E., Bronnenberg, B., Paap, R.,Wedel,M.,& Franses, P.H. (2010). Retrievingunobserved consideration sets from household panel data. Journal of MarketingResearch,47(1),63–74.

Venkatesh,V.,&Bala,H.(2008).Technologyacceptancemodel3andaresearchagendaoninterventions.DecisionSciences,39(2),273–315.

Verhoef,P.C.(2003).Understandingtheeffectofcustomerrelationshipmanagementeffortson customer retention and customer sharedevelopment.Journal ofMarketing, 67(4),30–45.

Verhoef, P. C. (2012). Klanten centraal in de bankensector. White paper, MonitoringCommissieCodeBanken,theNetherlands.

Verhoef, P. C., & Langerak, F. (2001). Possible determinants of consumers’ adoption ofelectronic grocery shopping in the Netherlands. Journal of Retailing and ConsumerServices,8(5),275–285.

Verhoef,P.C.,Antonides,G.,&DeHoog,A.N.(2004).Serviceencountersasasequenceofeventstheimportanceofpeakexperiences.JournalofServiceResearch,7(1),53–64.

Verhoef,P.C.,Franses,P.H.,&Hoekstra,J.C.(2002).Theeffectofrelationalconstructsoncustomer referrals and number of services purchased from a multiservice provider:Doesageof relationshipmatter?Journalof theAcademyofMarketingScience,30(3),202–216.

Verhoef, P. C., Langerak, F., & Donkers, B. (2007). Understanding brand and dealerretentioninthenewcarmarket:Themoderatingroleofbrandtier.JournalofRetailing,83(1),97–113.

WeberShandwick/KRCResearch(2012).BuyIt,TryIt,RateIt.RetrievedSeptember9,2015fromwww.webershandwick.com/uploads/news/files/ReviewsSurveyReportFINAL.pdf

Zeithaml,V.A.,Bolton,R.N.,Deighton,J.,Keiningham,T.L.,Lemon,K.N.,&Petersen,J.A.(2006).Forward-lookingfocuscanfirmshaveadaptiveforesight?JournalofServiceResearch,9(2),168–183.

2.2Value-to-firmmetrics

Introduction

Value-to-firmmetricsfocusonthevaluedeliveredbycustomerstofirms.Thesemetricsaretypicallybehavioralandarefrequentlyfinancialinnature.AswiththeV2Cmetrics,firmscanchoosefromaplethoraofV2Fmetrics.Someofthem,suchasbrandsalesandmarketshare, have been around for a long time, while others, such as customer lifetime value(CLV)arerelativelynew.BigdatadevelopmentshavecreatednewV2Fmetrics.Especiallyindigitalenvironments,manynewmetricshavebeendevelopedandarenowbeingusedindigitalmarketing.Thechallenge for firms ishow to interpretanduse thesenewmetricsandlinkthemtoexistingmetrics(Leeflang,Verhoef,Dahlström,&Freundt,2014).

In this in-depth chapterwe discuss several V2Fmetrics and theirmeaning.We againdivide metrics into market metrics, brand metrics and customer metrics. We will alsodiscuss new big data metrics, thereby specifically focusing on customer engagementmetricsandcustomerjourneymetrics.

Marketmetrics

Atthemarketlevel,firmswillmainlybeinterestedinmetricsthatshowtheattractivenessofthemarket.Thesemarketsizemetricsarerelevantforfirmswhentheymakestrategicmarketentrydecisionsandstrategicportfoliodecisions.BasedonthesetypesofanalysesacompanysuchasRoyalDutchPhillipscanmakestrategicdecisions: itdecided to retractfromtheelectronics-andlightningmarkettofocusonhealth.Marketmetricsarealsoveryimportant when considering new product sales. We therefore distinguish between twotypesofmarketmetrics:

1. Marketattractivenessmetrics2. Newproductsalesmetrics.

MarketAttractivenessMetrics

Figure2.2.1UKsmartphonesales

Source:AdaptedfromGfKRetail&Technology1

Typicalmarketmetricsbeingusedare:

Marketsize:Thetotalmarketdemandintermsofnumberof(potential)customers(i.e.targetpopulation),unitsor$sales(seeFigure2.2.1onUKsmartphonesales)Marketgrowth:TheannualgrowthofthemarketNumberofcompetitorsMarketconcentration:Thismetricmeasureswhetherthemarketisdominatedbyafewplayersorwhethermanyplayershaverelativeequalpositions in themarket.Theso-calledHerfindahl Indexisusedtomeasurethis. It isdefinedforaspecificmarketwithJfirmsas:

andcanvarybetween0and1,with1representingaveryconcentratedmarketwith

amonopolisthaving100%ofthemarket.

Oneproblemthatfirmsfaceishowtodefinethemarketandcompetition.Thisisbecomingrather difficult, because of market fragmentation. For example, strong retailers such asTesco not only sell fast-moving consumer goods, but also telecom subscriptions andfinancialservices.Similarly,newplayersareenteringthemarketinthisdigitalera.Google,forexample, canbecomeacompetitorofonline retailerswhen theyareactively startingGoogleShop;banksareconfrontedwiththegrowthofnewpaymentformssuchasPaypalandBitcoinsandarealsoworriedaboutGooglegettingintobanking.

Newproductsalesmetrics

For forecasting newproduct sales it is considered important to use severalmetrics (e.g.,Farris,Bendle,Pfeifer,&Reibstein,2006):

Trialrate:Thenumberoffirst-timenewproductusersasapercentageofthetargetpopulationRepeat volume: Number of repeat buyers multiplied by the number of productstheybuy in eachpurchase,multipliedby thenumberof times theypurchaseperperiodPenetrationrate:Numberofrepeatusersplusnumberofnewtrials,dividedbythemarketpopulation.

Based on these figures, volume projections of the market can be made. A very simpleequationforsalesvolumeisforexample:

Salesvolume=Penetrationrate×Purchasefrequency×Unitspurchased

Note that so far we describe the above metrics at the product level. However, similarmetricscanbecalculatedfor(new)brands.2

Brandmetrics

WeconsidertwotypesofV2Fbrandmetrics:

1. Brandmarketperformancemetrics2. Brandvaluationmetrics.

Brandmarketperformancemetrics

Brand market performance metrics focus on the actual performance of brands in themarket.Traditionalbrandmetricsarewellknownandareveryfrequentlymeasured:brandpenetration,brandsales,andmarketshare.Brandsalescanbemeasuredinternally.Brandpenetration andmarket share data aremore difficult tomeasure as information on thewhole market is required. In some markets, such as the FMCG (fast-moving consumergoods)market,thesemetricsaremeasureddailyorweeklyusingscanningtechnologiesbymarket research firms such as GfK, AC Nielsen, and IRI. In many other markets, forexample financial services and telecom, data on market shares are less frequentlymeasured.Onespecificissueisthatinthesemarkets,datacollectiononactualpurchasesisnotwellorganized,or is fairlydifficult toexecute.Forexample,whenmeasuringmarketsharesintheinsuranceindustry,customershavetoreportaccuratelyontheirownershipofinsurancesandthefirmstheypurchasedfrom.Itisquestionablewhetherthiscanbedoneinanaccurateway.

Nexttothesemoreaggregatebrandmarketperformancemetrics,firmsalsofrequentlymeasurebrandloyaltymetrics.Thesemetricscouldbebasedonstatedintentions,suchasbrand repurchase intentions. This can be measured using scales, such as for example aJusterscalewhere0=willabsolutelynotrepurchase,and10=willabsolutelyrepurchase(Juster, 1969).However,onecouldalsouse scales inwhichcustomershave todivide100points across brands in the market place—called a constant sum scale (Rust, Lemon, &Zeithaml, 2004).Thiswillmoreaccurately reflect the fact that customers frequentlybuymultiplebrandsandarefrequentlynotloyaltoonespecificbrand.Thispurchasebehaviorofmultiplebrandshasbeen referred toas “polygamous loyalty” (e.g.Dowling&Uncles,1997). Actual brand loyalty metrics concern the brand repurchase rate, which is thepercentageofcustomerspurchasingaspecificbrandthatwillrepurchasethebrandonthenext purchase occasion as well. It is important to note that a no-purchase on this nextpurchaseoccasioncanbe followedbya repurchaseof thebrand ina subsequentperiod.Usingthisinformation,aso-called“switchingmatrix”canbeformed(seeFigure2.2.2). Inthismatrix,oneobservestherepurchaseratesofabrandandtheswitchingprobabilitiestootherbrandsinthemarket.Inthisspecificexample,theswitchingprobabilityfrombrandA to B is 20%, whereas the probability of customers switching back to brand A in the

subsequentpurchaseoccasionis10%.Notably,thisswitchingmatrixnotonlycanbebasedonactualrepurchaserates,butcanalsousetheconstantsumscaleofthedivisionof100pointsacrossbrands(Rustetal.,2004).

Figure2.2.2Exampleofbrandswitchingmatrix

Withinmarketingsciencetherelationshipbetweenmarketshareandbrandrepurchaserateshasbeenstudied,andthegeneralfindingisthatthereisastrongrelationshipbetweenthem.Inparticular, theschoolaroundformerLondon-basedmarketingprofessorAndrewEhrenberg, with followers such as Byron Sharp, has aimed to demonstrate that thisempirical relationship can be shown in many markets. One of the most importantimplicationsof this is thatbrandswithahighmarket sharehaveahigh repurchase rateand that brands with low market shares have a low repurchase rate. There are someexceptions, such as a niche brand targeting a specific market segment, with loyalcustomers.OneoftheconclusionsEhrenbergandcolleaguesdrawisthatfirmsshouldnotbelievetoostronglyintheneedtocreateloyalcustomers.Theyarethereforeverycriticalon loyalty strategies, such as loyalty programs (Dowling & Uncles, 1997; Sharp, 2010).However,althoughonecoulddrawthisconclusion,theseanalysesonlyprovideacurrentstatus quo. There is sufficient evidence that loyalty strategies can create a higherrepurchaserateforbrands(e.g.Leenheer,Bijmolt,VanHeerde,&Smidts,2007).

Aswellaslookingattherevenuesideofbrands,brandinvestmentcanalsobemeasured.Thesebrandinvestmentscouldinvolve,forexample,advertisingcosts.Theresultsoftheseadvertising costs are measured with so-called “gross rating points” (GRPs). This iscalculated as the percentage of the reached target market multiplied by the exposurefrequencyoftheadvertisement.

Brandevaluationmetrics

Asbrandsareveryimportantassetsforfirms,therehasbeenstrongattentiononhowtofinancially evaluate these brands; in particular, financially oriented brand equity (BE)metricshavebeendeveloped.ThesemetricsdifferfromtheV2CBEmetricsthattypicallyfocusoncustomer-basedBEandfocusonlyonawarenessandattitudinalmeasures,suchasbrand preference. Probably the best known financially oriented BE metric is the onedevelopedbyInterbrand.EachyeartheydevelopthistocalculatetheBEofglobalbrandssuchasCoca-Cola,Apple,andBMWandpublishatop100.ForyearsCoca-Colawasthemost valuedbrand,untilApple tookover the first position in 2013, followedbyGoogle.

TheInterbrandmeasure3isbasedonthreepillars:financialperformanceofthebrand,roleofthebrandinpurchasedecisions,andbrandstrength,whichistheabilityofthebrandtocreate brand loyalty.4 The exact methodology of this metric is not fully shared andthereforerepresentsablackboxformanyresearchers.Beyondthesewell-knownmetrics,academicresearchershavealsoproposedsomemetricstoevaluateBE.Wediscusssomeofthesehere:thelistisbynomeansfullyexhaustive:

Brandequityshareholdervalueapproach:UsingthisapproachBEisbasedonthemarketvalueofthefirm(Simon&Sullivan,1993).Thismarketvalueisdecomposedintotangibleandintangibleassets.Thevalueof intangibleassetsconsistsofR&D(i.e.patents),valueof industry factors (e.g.monopolypositions),and thevalueofthebrand.Brand equity preference-based approach:With this approach, researchers aim tomeasure theabilityof thebrand toattractandkeepcustomers.Thiscanbedoneusingconjointanalysis(seeChapter4.1)inwhichtheutilityofanumberofproductvariants consisting of several attributes including the brand attribute is assessedusing several methodologies (e.g. choices between alternative products, productevaluations).AstrongerimpactofthebrandattributeontheproductutilityreflectshigherBE.Althoughusingadifferentmethodological approach, researchershavealsodevelopedBEmetricsusingdemandmodels.Inthisapproachthechoiceforaspecific brand over time ismodeled using product attributes andmarketingmixvariablesasexplanatoryvariables.Thevalueofthebrandconstantinthemodelisameasure forBEas it reflects thepower toattractandkeepcustomersbeyondtheeffectsofproductattributesandmarketing(Sriram,Balachander,&Kalwani,2007).Brand equity price premium approach: This BE metric focuses on the ability ofbrandstoaskahigherpricefromcustomers.Ameasureusedhereis“willingnesstopay.” A higher willingness to pay signals a higher BE. One problem with thismeasureisthatitissubjective,asitisbasedoncustomersurveysanditisdifficulttoquantifyfinancially.Anothermetricinvolvesmeasuringthepricepremiumofabrand. Ailawadi, Lehmann andNeslin (2003) propose the revenue premium as ameasure. This measure is a combination of brand unit sales and brand pricepremiumandmeasuresthepricedifferenceofthebrandwithagenericbrandandtheunitsalesdifferencewiththatgenericbrand(seeFigure2.2.3).Thechallengeistofindagoodgenericbrand. Intheirstudy,Ailawadietal. (2004)usetheprivatelabel(storebrand)asthegenericbrand.

Figure2.2.3Brandrevenuepremium

Source:AdaptedfromAilawadietal.,2004

Interestingly,studiesreportstrongcorrelationsofBEmetricswithproductmarketmetrics.Forexample,Ailawadietal.(2003)showthattheirrevenuepremiummeasureisrelativelystronglycorrelatedwithmeasureslikemarketshare.Onecouldprobablydebateabouttheadditional value of BE metrics over and above frequently used brand metrics, such asmarket share. One problem with these metrics is that they are frequently short-termoriented, while BE is more long-term oriented. Financial BE metrics should bediagnosticallyaboutabrand’slong-termhealth;thisgoesbeyondsalesandcouldconsiderfactorssuchastheattractivepowerofbrands,pricepremiumsandfinancialvalue.

Customermetrics

Customermetricsfocusonthebehaviorofindividualcustomersintheirrelationshipswiththefirm.Weadopttheso-called“relationshiplifecycle”asanunderlyingmodelofmanyofthesemetrics(Dwyer,Shurr,&Oh,1987).Inthisrelationshiplifecycle,thetrajectoryofanindividual customer moving from a prospective customer, to a customer, and finally adefectingcustomerisconsidered(seeFigure2.2.4).

Basedontherelationshiplifecycleconceptonecandistinguishthreetypesofmetrics:

1. Customer acquisition metrics, which focus on the first phase of attractingcustomers

2. Customer development metrics, which focus on how the relationship developsafteracquiringthecustomers

3. Customer value metrics, which consider the financial value of the customerduring the relationship and can involve both the acquisition phase andrelationshipdevelopmentphase.

Figure2.2.4Relationshiplifecycleconcept

Customeracquisitionmetrics

Customeracquisitionmetricsconsiderhowcustomersrespondtoacquisitionactions.Froma customer perspective we mainly consider individual customer acquisition techniques,suchasdirectmarketingefforts,telemarketing,etc.Onamoreaggregatelevel,firmswillbeinterestedinthenumberofcustomersacquiredthroughthesedifferentchannels.Fromacostperspectivetheywouldliketoknowtheacquisitioncosts(percustomer).Importantly,

wealsosuggestconsideringthevalueoftheacquiredcustomers.Firmsconstantlyfallintothe trap of acquiring many customers with attractive price offers who are likely to beunprofitable and will switch frequently after one year (Lewis, 2006). When evaluatingacquisitioncampaignstwospecificmetricsarefrequentlyconsidered:

Responserate: thenumberofcustomersrespondingtoacampaigndividedbythenumberoftotalcustomersapproachedwithacampaignConversion rate: the number of customers acquired divided by the number ofcustomersrespondingtoacampaign.

Feld, Frenzen, Krafft, Peters, & Verhoef (2013) discuss some intermediate metrics toevaluate campaigns in addition to the above-mentioned metrics. They report on theopeningrateofadirectmailingandthekeepingrateofthedirectmailing.

Customerdevelopmentmetrics

Customerdevelopmentmetricsconcernmanydimensionsofrelationshipdevelopment(e.g.Bolton,Lemon,&Verhoef,2004).Specifically,weconsiderthreetypesofmetrics:

1. Relationshipcontinuationorlengthmetrics2. Relationshipexpansionmetrics3. Relationshipcostsandriskmetrics.

Relationshiplengthmetrics

Thesemetricsfocusonthecontinuationoftherelationship.Specificmetricsconcern:

Churnorcustomerdefection:ThepercentageofcustomersquittingtherelationshipwiththefirmCustomerretention:Thepercentageofcustomerscontinuingtherelationshipwiththefirm(1–customerchurn/defection)Customer lifetime / relationship length: The (expected) length of the relationshipbetweenthecustomerandthefirmPurchasefrequency:ThenumberoftimesacustomerpurchasesfromacompanyinaspecifictimeperiodRecency:Timesincelastpurchase.

One specific metric you could consider here is the win-back percentage of defectedcustomers. Win-back metrics become prevalent when firms actively approach churnedcustomerstobecomecustomersagain.

Relationshipexpansionmetrics

Relationshipexpansionconcernsmetricsthatfocusonthegrowthoftherelationship.Thisexpansioncaninvolvemultiplemetrics:

Averagenumberofproducts/servicessoldpercustomerCross-buyingrate:ThepercentageofcustomerspurchasingadditionalproductsorservicesfromacompanyUpgradingrate:Thepercentageofcustomersupgradingtheirproductsorservicestoahigherlevel(e.g.upgradedservicecontract)Adoption rate: The percentage of customers purchasing the newly introducedproductorserviceCustomershare:Thenumberofproductsorservicespurchasedbyacustomerfromafirmdividedbythenumberofproductsorservicespurchasedbyacustomer inthatspecificproductcategoryShareofwallet:Themoneyspentbyacustomerata firmdividedby themoneyspentbyacustomerinthatspecificproductcategory.

The first four measures can be gained from the customer database. The share metricsprovide information on the relative loyalty position of a firm for a specific customer. Inordertocalculatethesemetricsadditionaldataonthepurchasebehaviorofcustomersinaspecificcategoryisrequired(Verhoef,2003).Highercross-buyingratesshouldusuallyalsobe reflected inhigher values for the sharemetrics. Importantly, recent research suggeststhatcross-buyingcanindeedhavepositiveprofitconsequences,butnotallcustomerswithlargecross-buyingpercentagesareprofitable(Shah,Kumar,Qu,&Chen,2012).

Relationshipcostsandriskmetrics

During therelationship, firmsalso invest in thecustomerrelationship. Investments focusondevelopingtherelationshipandcan,forexample,includethecostsofaloyaltyprogram.WediscussthesecostsinmoredepthinthenextsectiononCLV.Costsmaymainlyinvolvethecosts to servean individual. Importantly, thesecostsmayvaryconsiderablybetweencustomers.Onespecificriskthatfirmsfrequentlyfaceisthatcustomersmightnotpay.Inthis regardmeasuring the debt risk of the customer base and individual customers is ofessential importance, especially for firms delivering products or services before thepayment is received (e.g. utilities, telecom). Debt risk can be modeled and predicted(L’Hoest-Snoeck,VanNierop,&Verhoef,2015).Forretailersthenumberofproductreturnshasalsobecomeanimportantcost-relatedmetric.Returnsoccurwhencustomersorderaproduct that does not deliver what they actually hoped for. Frequently, customers canreturntheseproductsforfreeandfirmsareobligedtoreceivethemback—andhavetopay

backthepricepaid.Reducingreturnrateshasbecomeoneofthetopprioritiesforonlineretailers, such as Zalando, in order to improve profitability. Overall, Shah et al. (2012)suggest that firms should consider metrics that reveal so-called “adverse behavior” ofcustomers, which may involve debt risk, large service costs, and product returns.Interestingly, customers showapattern in these adverse behaviors and thus in targetingpoliciesfirmscanchoosenottoofferthesecustomersattractiveoffers(Shahetal.,2012).

Customervaluemetrics

The metrics discussed so far focus on actual behavior during the relationship lifecycle.Customervaluemetricsconsidertheresultingfinancialvalueofcustomers.Valuemetricscanbedividedbasedontheforward-lookingnatureofthesemetrics.Non-forward-lookingvaluemetrics focusonthecurrentstatusofacustomerandinvolvecustomerrevenueorthemonetaryvalueofacustomer,customermargin,andcustomerprofitability.Forward-looking metrics consider the expected value of customers in the future. The mostprominentmetricinthisregardiscustomerlifetimevalue.Thismetrichasgainedsomuchattentionintheliterature—andhasbeenacceptedinpractice—thatwedevoteanextensivediscussionofthismetricinthischapter.

Customerlifetimevalue

Customerlifetimevalue(CLV)isametricthathastakenoffwiththestrongdevelopmentofcustomerrelationshipmanagement(CRM).Thismetricistypicallyusedtoevaluatethevalueofcustomers.CLVisfrequentlydefinedas“thepresentvalueofthefuturecashflowsattributed to the customer during his/her entire relationshipwith the company” (Farris,Bendle,Pfeifer,&Reibstein,2010).Inthisdefinition,CLVassessesthetotalvaluedeliveredby a customer. However, firms are frequently more interested in the future value ofcustomers. CLV can then be considered as a forward-looking customer centric metric,basedonassumptionsandpredictions,andcanbedefinedasthenetpresentexpectedvalueof a customer. In otherwords, value created in the past is not taken into accountwhencalculatingthevalueofacustomerorcustomerbase.However,pastcustomervaluecanbea predictor of future value (e.g. Donkers, Verhoef, & De Jong, 2007). CLV is a veryimportant metric, because (when calculated properly) it can be the link between themarketing and the financial department and create the commonplatform to bring theseworlds together. Furthermore, there is sufficient evidence to show that,when calculatedproperly,CLVcanbeagoodindicatorforfirmvaluation(e.g.Gupta,Lehmann,&Stuart,2004),especially inthose industries/companieswherecustomersarethebiggestassetandmarketsareratherstable.AverysimpledefinitionofCLVforcustomeriattimet=0andwhere“d”isthediscountrateis:

Inthisdefinitionweassumethateachyearaspecificmarginisearnedpercustomer.ThesemarginsaresummedovertimeuntilachosenendpointT.Typically,periodsof3–5yearsareusedforCLV-calculations(Rust,Zeithaml,&Lemon,2000).Adiscountrateisusedtomakethefutureearningspresent.Thisdiscountrateissetincooperationwiththefinancedepartment.AhighdiscountrateimpliesthatfutureearningscontributelesstoCLVandmay signal that firms value these future earnings less. This implies a more short-termorientation.

CLVanditscomponents

ForaproperCLVcalculation,itiscrucialthatinbuildingtheunderlyingCLVmodel,theassumptions,components,andoutcomesareadoptedandacceptedbyboththemarketingandthefinancedepartment.BytheCLVmodelwemeanthedifferentcomponentsofCLV,howtheyarecalculatedandhowtheyareinterrelated.ThemaincomponentsofCLV(seealsoFigure2.2.5)are:

1. MarginorEBITDA

2. Expectedlifetime3. Onetimeinvestments.

Margin

MarginorEBITDAistheresultoftherevenuesgeneratedpercustomerminusoperationalcosts.Forpurecalculations, especially incapital intensive industries,depreciation shouldalsobetakenintoaccount,butwewillskipoverthistopreventunnecessarycomplexity.Usually for the revenuepartweonly look at thebilled revenue,meaning thatdiscountsgiven have been deducted from the gross revenue. Especially when discounts aresubstantial and can differ between customers, there can be a considerable differencebetweengross revenue andnet revenue.Using the charged amount also implies thatwewillusethefrequencyofbilling(forexamplemonthly)inacontractualsettingasthetimeunittobeusedinourCLVmodel.

Figure2.2.5TheCLVmodel:theelementsofcustomerlifetimevalue

SincewearecalculatingtheCLV, foramulti-productorganizationweshouldsumtherevenueperproductgeneratedby thecustomer.Expecting thecustomer to staywith theorganizationformanybillingcycles, ideallymanyyears, itwillbenecessarytocalculatethe net present value of future revenues (see Equation 2.2.2). In otherwords, a euro ordollar collectednextyear isworth less than theeuro/dollarof thisyear.Wewilldiscussthis further when explaining lifetime calculation. To obtain the margin from revenues,costshavetobedeductedfromrevenues.Wedistinguishtwotypesofcosts:

Costsofgoodssold(COGS):thecostdirectlyassociatedwiththeproductusage(forexampleminutescalledforatelcocompany)ormanufacturingoftheproduct;Overheadcosts:thecostsoftheorganizationnotdirectlyrelatedtoproductusageormanufacturing,forexampletheITdepartment,thecallcenter,realestateetc.

It is important that the total costs of the organization are allocated to one of these two

buckets, to preventmissing certain costs, but also to prevent double counting. Since theelementsofCOGSaredirectly linked toa singleproduct they canbe easily allocated toindividual customers buying and/or using different products. Calculating the COGS isnormallypartoftypicalcostpricecalculationseveryorganizationshouldhave.TobuildagoodCLVmodel, theCOGS ideally shouldbedecomposedasmuchas possible, to copewithdifferentcostpricespertypeofproduct/service.Forexample,foratelcocompanythecostpriceofa textmessagemightbe lower than thecostpriceofaminutecalledandaminute called can be further broken down into aminute called on the company’s ownnetworkoronthenetworkofanotherorganization.HavingtheCOGSpercustomer,basedon total product usage (built up from the different components of product usage) canalready result in large differentiation within the customer base: for example, customerspayingthesamefixedsubscriptionfee,withdifferencesintypeandintensityofusage,canresultincompletelydifferentmargins.

Fortheoperationalcostcenters,thefirststepistoidentifythelargecostbucketsinthecompany’sP&Laccount.Forthelargestcostbucketsourchallengeistoidentifythedriversofcustomerbehaviorofthesecosts,inordertodefinetherightallocationmethodologyofthesecosts.Let’stakeforexampletheyearlytotalcostsofthecallcenter(humanresources,tooling, housing, infrastructure, etc.). Obviously the costs are driven by the number ofcustomerscalling,ormorespecificallybythenumberofcallsandperhapsthedurationpercall. These numbers can be used to allocate the costs of the call center to individualcustomers.However, part of these total costs should be allocated to all customers, sinceevenifnotonecustomercalls(notrealisticbutlet’sassume),theorganizationstillwouldhaveacallcenterwithallitsfixedcostsinplace.That’swhynon-callingcustomersshouldalsobeallocatedpartof the call center costs.Typically, this isdonebyanalyzingwhichpartofthetotalcostsarefixedanddividethesebythetotalnumberofcustomers,andusethecustomerbehaviordriversforallocatingallothercosts.Forthesmallercostbuckets,theeasiestway is to just spread thesecostsacrossall customers. InBox2.2.1weprovideanexampleofallocationofcoststoguideanalystsfurther.

Box2.2.1Exampleofcostallocationoptionsfromsimpletocomplex

Totalyearlycostofcallcenteris10millioneuros.Totalnumberofcustomersis1million.

Inputonpossibledriversofcostallocationtoindividualcustomers:

Totalnumberofcallsis2million.Totalnumberofcustomerswithatleastonecallis500k.Averagenumberofcallspercallingcustomeristhus4calls.

Thehighestnumberofcallspercustomeris25.20%ofthecallsareaboutbilling,60%aretechnicalquestions,20%arecontractortransactionrelated.Billing calls take on average 2 minutes, technical calls 15 minutes andcontract/transactioncalls10minutes.20%ofcostsarefixed,80%arevariable.

Differentoptionsforcostallocation,fromsimpletocomplex:

Option1.FlatallocationDividethetotalcostsofcallsbythetotalnumberofcustomers:10million/1millionandallocate10eurospercustomer.

Option2.FixedandvariablesplitDivide the 80% variable part of the costs by the number of calling customers: 8million/500k=16eurosandadd20%ofthefixedcosts(2million)dividedbythetotalnumber of customers = 2 million /1 million = 2 euros. This means that a callingcustomergets18eurosofcostsallocatedandanon-callingcustomer2euros.

Option3.CallratioAs inoption2,anon-callingcustomersgets2eurosofcostsallocated,butnowweallocatetheother80%ofthecostsbylookingatthenumberofcallspercustomer.Percall thismeans8millionvariablecosts / 2millioncalls=4eurospercall. Inotherwords, a customerwith25 callswill get (25x4)+2 eurosof costs allocated=102euros,whileonaverageacallingcustomerwillstillget(4×4)+2=18eurosofcostsallocated.

Option4.TypeofcallThe total minutes called are (400,000 × 2) + (1,200,000 × 15) + (400,000 × 10) =22,800,000.Thismeansthateveryminutecalledcosts8million/22.8million=0.35europerminute.Soabillingcallcosts0.35×2minutes=0.70euro,atechnicalcall0.35×15=5.27andatransactioncall0.35×10=3.50euros.Thismeansthatacallingcustomerwithatechnicalcallandatransactioncallwillgetintotal5.27+3.50+2.00=10.77eurosallocated.

Otheroptionscanbebuiltup,forexamplebyafurtherspecificationofthetypicalcostsper typeof call, orby taking intoaccount theactualminutesnecessary for acall.Sotheaboveisnotexhaustive,butshowshowfurthercostdifferentiationbasedon the drivers and type of customer behavior and characteristics can help in costallocation.Inthisthedatabaseandthedatabaseanalysesbythedata-analystsplayanimportantrole.

Lifetime

Lifetime is the metric that indicates the total (expected) remaining duration of therelationship.Thisisthemetricwiththehighestimpact(duetothemultiplieronthemarginitdelivers);however,itisalsothemostcomplextocalculate.Afirstdistinctionthatshouldbemadeisbetweenthetwodifferenttypesofrelationships(Fader&Hardie,2010):

1. Thecontractualsetting:thecustomercommitsforacertainperiodtouseaserviceor product and cannot leavewithout good reasons (death,moving abroad, etc.)within this period. This type of setting is often seen in utilities, telecoms, andinsurance. In this setting theminimumlifetime is the remainingdurationof thecontract.Whenthecontractendsthereisprobabilitythatthecustomercontinuesandrenewsthecontract.Thiscontinuationisobserved.Usuallythecustomerhastodosomething(e.g.sendaletter,fillinaform,call)toendthecontract.

2. Thenon-contractualsetting:thecustomerhasnocontractualcommitmenttotheorganizationandcandecideatanymomenttoswitchorstopusingtheproductorserviceandhasnoobligationtocommunicatethistotheorganization.Thiskindof behavior is very typical in (online) retailing andwith fast-moving consumergoods.Inthissettinglifetimeisbasedontheprobabilityoffutureusageorbuyingoftheproductorservice.Typicallyswitchingisnotobservedinnon-contractualsettings, as customers donothave to end a contract.They just stopbuying.Assuchoneneverknowswhenarelationshipends.

Therearedifferentdriversthatinfluencecustomerlifetime,suchasusage,pastrelationshipduration,socio-demographics,andproductsused.Thesedriverscanbeincludedinmodelsto predict lifetime. Typically, differentmodels are used for contractual settings than fornon-contractual settings. For contractual settings analysts tend to use discrete choicemodels, such as the logit-model (e.g., Donkers et al., 2007) (see Chapter 4.1). For non-contractual settingsdurationmodelsornegativebinomialdistribution (NBD)modelsareused(Fader&Hardie,2010).

Investments

Weconsider as investments all net organizational expenses that are spent fromamulti-periodperspectiveonthecustomerrelationshipandthatarenotcoveredbytheCOGSoroperational cost centersand that result in “out-of-pocket” for theorganization.One-timerevenuesaresubtractedfromthetotaloftheseexpenses(forexample,thefeecollectedforconnectingasubscriptionforamobiletelco).Componentsoftheseinvestmentscanbe,forexample:

AsubsidyondeliveredhardwareGadgets/premiumsMarketing campaigns above the line (ATL) and below the line (BTL) (where the

costsofthecampaignsareallocatedtotheconversion)Channelcosts,suchasfeestobepaidtointermediaries…andsoon.

Theseinvestmentsaremadeinthebeginningoftherelationship(acquisitioninvestments)to composeanappealingoffer to thenewpotential client, or at thepotential endof therelationship (for example, as the contract is expiring), in order to renew the contract(retention investments).These investments in thecustomer relationshipdiffer largelyperindustryandarebasedontheexpectedrevenuestreamduringtherelationship,thecostsofthe relationship, and the expected length of the relationship. Since in the contractualsettingtheminimumrevenuesandlengthoftherelationshipareknown,creatingalowerlevel of uncertainty, this will often result in higher upfront investments. In certainindustriessuchasinsurance,telecoms,orutilitiesthismightresultininvestmentpernewcontractofseveralhundredsofeuros/dollars.Inthenon-contractualsettingtheseupfrontinvestments are usually low, due the much higher uncertainty of future revenues andlifetime.Evensmallinvestmentsaredifficulttoearnback.Forexample,smallinvestmentstoacquirenewcustomersfortheUSonlinegrocerystart-upRetailRelaywerehigherthantheexpectedCLVofanewcustomer(Venkatesan,Farris,&Wilcox,2014).

Aswehaveseenearlier inthischapterwhendiscussingthecalculationofmarginandlifetimeperindividualcustomerandtheoutcomeofthecombinedmetrics(whichwecallgrossCLV),onecanexpectbigdifferencesbetweenindividualcustomers(seeFigure2.2.6).ThismeansthatanorganizationthatistryingtobalancegrossCLVandinvestmentscanuseinvestmentsandespeciallythedifferentiationintheinvestmentsasawaytooptimizeitscommercialefforts.

Calculatingcustomerlifetimevalue

ForcalculatingtheactualCLV,basedonourCLVmodelanditscomponents,wewillneedaformulatoincorporateallelementsinordertocalculateaCLVperindividualcustomerthat canbe expressed as a euro/dollar amount.At aminimum level, firmsneed tohavedataoncustomermarginsandexpectedlifetime.Beyondthat,especiallywhencalculatingthe value of a new customer, data on investments in acquisition or retention should betakenintoaccount.ExtensiveliteratureonCLVhasproposedseveralmoreextendedCLVmodels(e.g.,Berger&Nasr,1999;Venkatesan&Kumar,2004;L’Hoest-Snoecketal.,2015).We specifically discuss a more extensive version of the simple equation in which weincludetheretentionrate(r)toaccountfortheexpectedlifetimeofindividualcustomers,therebytakingintoaccountthepossibilitythat

Figure2.2.6ExampleofgrossCLVdistributionperdecile

earnings from customers may drop because customers churn and that firms invest inacquisition:

An easy way to calculate CLV is to take T to infinity. After some mathematicalcomputationsthefollowingsimpleformulaisachievedforthecalculationoftheCLVforcustomeri:

Eachoftheabovecomponentmodelscanbeusedtoforecasttheindividualretentionratesand margins. For example, retention rates can be predicted using a logistic regressionmodel.Moreextensiveversionsoftheabovemodelscanbedeveloped;onewaymightbetoconsidermargingrowththrough,forexample,reducedcostsorincreasedrevenuesfromcross-buyingorupgrading.WerefertoBergerandNasr(1999)foranextensivediscussionof these basic models. Donkers et al. (2007) also discuss many models that differ incomplexity.Interestingly,theyshowthatintheirsetting,thesimpleforecastruleoftoday’scustomer profit is next years’ profit has the strongest predictive performance. In oursubsequentdiscussionoftheCLVmetrics,wewilltakeamorepragmaticperspectiveanddiscusshowfirmscaneffectivelyimplementCLVasanimportantmetric.

GettingstartedwithCLV:Bepragmatic

InbuildingaCLVmodel,oneshouldrealizethatitisquiteimpossibletobuildafullblownCLVmodelfromscratch.Inrealityaphasedapproachismuchmorerealisticandcanhelpin processing new insights on the dynamics of the CLV model. Defining the different

phases in a pragmaticway on a project basis can help the organization gain experiencewith CLV at an early stage and reduce the probability of failure. Starting with a toocomplexmodel in the beginning can result in disappointment, the loss of organizationalacceptance, inaccurate results, and delays in delivery. It is good to know that there aremanypragmaticapproachesandrulesofthumbavailabletorealizeafirstCLVmodelforquickresults.

Let’sstartwiththeCLVformula:

The first component, the (monthly) margin (revenues minus costs), could besimplifiedbyjusttakingapercentageoftherevenue.Thiscouldbefurtherrefinedbyspecifyingthismarginpercentageperproduct,assumingthatperproductacostallocationmodel delivers this percentage, differentiating this per product.Anextstep in margin calculation could be to take the largest cost buckets and try toallocatethembasedonconsumerbehavior.Agoodproxyofthelifetimeofthecustomer,basedontheretentionrate(r)intheformula above for a contractual setting, can be obtained by assuming that thelifetimeequals thedurationof the first contract. Inanon-contractual settingonecould take thebuying frequencyperyear (orother timeframe)as an indicatoroflifetime.Itmightbeclearthattheaboveproxiesarestillaratherroughproxyofthereallifetime,withonlyalittledifferentiationpercustomer.Soanextstepcouldbeto build a first basic churn prediction model to allocate customers into churnbuckets and assigning a churn (in a contractual setting) or inactivity (in a non-contractual)settingprobabilitypersegment.Theeasiestwaytoestimatetheinvestmentsinacquisitionorretentionistotakethetotal investments and divide this by the total number of transactions for new orrenewing customers. This could be refined by splitting these total costs inacquisition and retention investment and dividing this by either the number ofacquisitions or the number of retentions. The next step could be to specify theinvestmentsperchanneland/orproduct.

In Chapter 6, one of the cases we will discuss will be a CLV case for a large energycompanythattriedtomodelCLVoverthecustomerbase,toassesstheimportanceofeveryelement of theCLV formula.We also suggest taking a look at the numerous calculatorsthatcanbeusedasaguideinbuildingyourownCLVmodel.SpecificallywewouldliketomentionthecalculatordevelopedbyHarvardBusinessSchool.5EventhebasicversionisaveryniceexampleofhowtocalculateCLV.

Customerequity

CustomerequityiscloselyrelatedtoCLV.Themetric,assuccessfullyproposedbyRustetal.(2000),isthesummationofallCLVsofcurrentandfuturecustomersofafirm.AssuchitisbroaderthanCLVbecauseitconsidersallcustomersandalsofuturecustomers.Hence,itconsidersboththevalueofexistingrelationshipsresultingfromcustomerloyaltyandtheabilityofthefirmtoattractnewvaluablecustomers.Toachievethis,Rustetal.(2000)useaswitching matrix approach, in which customers can switch between suppliers. Insubsequent work Rust et al. (2004) show that firms can calculate the consequences ofinvestmentsindriversthatincreasethevaluedeliveredtocustomers(reflectedincustomerperceptions on customer equity), by considering acquisition and retention consequencesand subsequent effects on CLV and customer equity. By comparing the investments invaluecreation(e.g.increasinglegspaceinairplanes)withthecustomerequitychanges,amarketingROIcanbecalculated(seeFigure2.2.7).

Newbigdatametrics

Asaresultofthedevelopmentofbigdata,wecanconsidertwoimportantnewareasthatrequireadditionalmetrics:

CustomerengagementDigitalcustomerjourney.

Customerengagement

Theincreasingpresenceofsocialmediahasstirredupattentionforthenon-transactionalbehavior of customers. Customers not only add valuewith their purchase behavior, butmayalsoaddvaluebysharing theirexperiencesonline, influencingothercustomersandprovidinginputthroughco-creation.Thisnon-transactionalbehaviorisfrequentlyreferredto as “customer engagement behavior” (vanDoorn et al., 2010; Hoyer, Chandy,Dorotic,Krafft,&Singh,2010).Thisbehaviorresultsinapossibleneedforadditionalmetrics,suchas:

ThenumberofreferralspercustomerThenumberofrelationshipsofacustomerwithothercustomersThe number of ideas (e.g. new products, service improvements) of customersprovidedtoafirmThe influentialpowerof customers (i.e.measuredbyopinion leadershipor socialnetworkvariables;seeforexampleRisselada,Verhoef,&Bijmolt,2015).

Oneproblemwiththesemetricsisthattheyarefrequentlydifficulttomeasureasitmayinvolvesocialnetworkinformationand/orself-reportsoninfluence.Wewillreflectonthisinmoredepthinthe in-depthChapter4.2onanalytics,wherewediscusssocialnetworkanalytics.Firmshave,however,beenabletocollectdataintheirdatabasesonreferralsandpotentially onnumbers of ideas (e. g.throughmeasuring complaints). The importance ofcustomer engagement also implies an extension of theCLV concept. SpecificallyKumar,Donkers, Venkatesan,Wiesel, and Tillmanns (2010) introduced the concept of “customerengagement value.” Within this concept three new additional value components areintroduced: customer referral value (CRV), customer influencevalue (CIV) and customerknowledgevalue(CKV),asshowninFigure2.2.8.

Figure2.2.7CustomerequityROImodel

Source:AdaptedfromRust,LemonandZeithaml(2004)

Figure2.2.8Customerengagementvalue:ExtendingCLV

Source:Kumaretal.(2010)

Kumaretal. (2010)distinguishbetweencustomer-to-customer (C2C)andcustomer-to-firm(C2F)values.CRVandCIVareC2Cvalues,whereasCLVandCKVareC2Fvalues.CRV has been operationalized andmeasured using actual referral behavior in work byKumarandcolleagues(Kumaretal.,2010;Kumar,Bhaskaran,Mirchandani,&Shah,2013).Interestingly,theyshowthatcustomerswithamediumCLVhavethehighestCRV.CIVismore difficult to measure, because of the need for network data. Kumar et al. (2013)measuredCIVforthesocialmediacampaignofanIndianice-creamretailer,HokeyPokey.TheycalculatedbothCLVandCIVandsummingthesemetricstheycalculatedtheROIofthesocialmediacampaign.

Customerjourneymetrics:Pathtopurchase

Thedigitalrevolutionhasledtoanewomni-channelenvironment(e.g.Verhoef,Kannan,

&Inman,2015).Customersarenowdevelopingtheirownpathtopurchase.Theybrowseand searchonline, switchbetweenofflineandonlinechannels,usemultipledevices, etc.andarestillbeing influencedby traditionaladvertising. In sum,differentcustomers facebrandsatdifferenttouchpointsthatmayaffectdifferentcustomersindifferentways(e.g.Baxendale,Macdonald,&Wilson, 2015). This new development also results in amix ofnew brand and customermetrics. At the brand levelwemay, for example, observe thenumber of click throughs on a banner ad. For customers, firmsmay observe conversionrates.However,forprospectivecustomersthesemetricsarelesseasytomeasure.Forfirmsitisstillimportanttounderstandthispathtopurchasesandtomeasurespecificoutcomesalongthispathtopurchase,forbothcustomersandprospectivecustomers.(e.g.,Verhoefetal.,2015;Li&Kannan,2014).Thisresultsinasetofnew,differentdigitalmetrics:

Numberofwebsitevisits:whenpayingotherwebsitestoshowyouradvertisementand/or banner the associated cost is measured against the costs per 1000views/eyeballs(CPM=Costpermille)Click through rates: The percentage of customers viewing an online ad, searchengineoutcome,etc. clickingon thead tovisit the referredwebsite; the financialmetric to represent the associated costs when, for example, showing an onlineadvertisementviaGoogle—costperclick(CPC)Purchaseconversionrate:Thenumberofpurchasesafterawebsitevisitdividedbythe number of website visits. The financial metric to show the costs ofadvertisementsoronlineactivitiesiscostperorder/transaction(CPO)Averageordersize:TheaverageordersizeofeachpurchaseCostsofeachuniquetouchpointChannelswitching:Themigrationpatternsofcustomertootherchannels(especiallyrelevantwhenfirmsaimtomigratecustomerstolow-costschannels(e.g.,Trampe,Konuş,&Verhoef,2014;Gensler,Verhoef&Böhm,2012)Research shopping percentage: Percentage of customers searching in one channeland purchasing in another channel (e.g. Verhoef,Neslin,&Vroomen, 2007). Onespecific formof this is the showroomingpercentage,where the searchchannel isthe store and the purchase channel is online (Rapp, Baker, Bachrach, Ogilvie, &Beitelspacher,2015).

WewilldiscusscustomerjourneyanalyticsandpathstopurchasemodelsinmoredepthinChapter4.2.Onespecificconcernmanagersfaceistolinkinvestmentsindifferent(digital)acquisition channels to purchase outcomes. This is not so trivial, as customers may beinfluenced bymultiple channels or touchpoints in their purchase decisions and specificchannels(e.g.searchengines)arebydefinitionclosertothepurchasedecisionthanotherchannels (e.g. advertising). Firms require strong attribution models to quantify thecontributionofeverychannel.

MarketingROI

OnefinalmetricthatfirmsareinterestedinistheROIofmarketinginvestments.Wehavealreadybrieflymentionedthisinourdiscussionofthecustomerequitymetric,sinceROIisdirectlyrelatedtocurrentandfutureCLV.TheROIonmarketinginvestmentsiscalculatedastheadditionalCLVdividedbymarketinginvestment.

IdeallymarketingROIshouldcoverallpossiblemarketingactivitiesthatcanbedeployedand not just the (direct) marketing activities addressing new and existing customers.Furthermore,themeasuretocalculateROIshouldnotbebasedonsalesvolume,butideallyon created CLV, and it should also take into account the extent to which marketingactivitiescontributetoV2Cmetrics(seeFigure2.2.9).

Figure2.2.9ExampleofROIcalculation

In fact, amarketingROI calculation should also encompass, for example, investmentsmadeinATLcampaigns,specificbrandactivities,andothermarketingmixelements.ButtherearefewcasesthatsuccessfullyshowmarketingROIasaholistic(i.e.analyzingCLVandV2Ceffectsofthewholemarketingmix)approach.

Thereareseveralreasonswhythisisnotcommonpractice:

The relationship between, for example, an ATL campaign and the extra CLVcreated is very difficult to assess. This is because there are all kinds of possibleinterferencesonthefinaleffect(ifany)onCLV.However,thereisawholeworldofstudies and projects going on trying tomakemarketingmixmodels in order toquantifytheserelations.Integratingalldatasources,inordertomeasurealltheeffortsbeingdeployedandtheireffectonintermediateKPIslikebrandawareness,marketshareetc.insteadofmeasuring the effect onCLV, is very complex.This challenge of data integrationandhowtodealwithitwillbediscussedindepthinChapter3.1.Furthermore,CLVmight raise some discussionswithin the organization, becausesomeof theelementsof theCLVcalculationareconsideredeitherablackboxorare based on assumptions (like, for example, expected lifetime) that are not

indisputable.InthesecasesorganizationsmightchoosetofallbackonsalesorsalesrevenuesinsteadofCLVformarketingROIcalculations.Thereisalackofanorganization-wideacceptedsegmentationofcustomersandthemarket that canbe identified inall sources in scope.Notonlydoes this limit thepossibilitiesofdataintegration,butitalsolimitstheactionabilityofinitiativestoimprovethemarketingROI.

Conclusions

In this chapterwe have discussedV2Fmetrics at both themarket, brand, and customerlevel. It is clear that there is a plethora of metrics from which firms can choose. Thisactuallyholds forbothV2CandV2Fmetrics.Wehaveaimed toprovideanoverviewofrather frequentlyusedmetrics,butacknowledgethatonecaneasilycomeupwithmanyotherrelevant(orirrelevant)metrics.ForinterestedreaderswerefertoFarrisetal.(2006),who provide a very extensive discussion of multiple metrics for multiple areas ofmarketing. Importantly, we also discussed some new big data metrics. Specifically, wefocusedoncustomerengagementmetricsandcustomerjourneymetrics,aswebelievethatmajor developments are occurring here. A clear understanding of thesemetrics will berequired to actually implement big data analytics to show the ROI of (additional)marketinginvestments.

Notes

1Seehttp://followthatapp.co.uk/2011/08/10/uk-smartphone-sales-statistics(accessedSeptember17,2015).

2WerefertoFarrisetal.(2006)foradetaileddiscussiononsomeofthesemetricsandhowtheycanbeusedtocalculate

salesvolumes.InourChapter4.1wealsodiscusssomemarketforecastingtechniques.

3 Other agencies have also developed BE metrics. The Interbrand metric is, however, best known and the most

influential,andwethereforehaveincludedthismetricinthischapterandignoredsomeothermetricsofcommercial

agencies.

4Seewww.interbrand.com(accessedSeptember14,2015).

5Seehttps://cb.hbsp.harvard.edu/cbmp/resources/marketing/multimedia/flashtools/cltv/index.html(accessedSeptember

14,2015).

References

Ailawadi,K.L.,Lehmann,D.R.,&Neslin,S.A.(2003).Revenuepremiumasanoutcomemeasureofbrandequity.JournalofMarketing,67(4),1–17.

Baxendale, S., Macdonald, E. K., & Wilson, H. N. (2015). The impact of differenttouchpointsonbrandconsideration.JournalofRetailing,91(2),235–253.

Berger, P. D., & Nasr, N. I. (1999). Customer lifetime value: Marketing models andapplications.JournalofInteractiveMarketing,12(1),17–30.

Bolton, R. N., Lemon, K. N., & Verhoef, P. C. (2004). The theoretical underpinnings ofcustomer asset management: A framework and propositions for future research.JournaloftheAcademyofMarketingScience,32(3),271–292.

Donkers,B.,Verhoef,P.C.,&DeJong,M.G.(2007).ModelingCLV:Atestofcompetingmodelsintheinsuranceindustry.QuantitativeMarketingandEconomics,5,163–190.

Dowling, G. R., & Uncles,M. (1997). Do customer loyalty programs reallywork? SloanManagementReview,38(2),71–82.

Fader,P.S.,&Hardie,B.G.(2010).Customer-basevaluationinacontractualsetting:Theperilsofignoringheterogeneity.MarketingScience,29(1),85–93.

Farris,P.W.,Bendle,N.T.,Pfeifer,P.E.,&Reibstein,D.J.(2006).Marketingmetrics:Fifty+metricseverymarketershouldknow.Philedelphia:WhartonSchoolPublishing.

Farris,P.W.,Bendle,N.T.,Pfeifer,P.E.,&Reibstein,D.J.(2010).Marketingmetrics:TheDefinitiveguidetomeasuringmarketingperformance.UpperSaddleRiver,NJ:PearsonEducation.

Feld,S.,Frenzen,H.,Krafft,M.,Peters,K.,&Verhoef,P.C.(2013).Theeffectsofmailingdesign characteristics on directmail campaign performance. International Journal ofResearchinMarketing,30(2),143–159.

Gensler, S., Verhoef, P. C., & Böhm,M. (2012). Understanding consumers’ multichannelchoicesacrossthedifferentstagesofthebuyingprocess.MarketingLetters,23(4),987–1003.

Gupta,S.,Lehmann,D.R.,&Stuart,J.A.(2004).ValuingCustomers.JournalofMarketingResearch,4(1),7.

Hoyer, W. D., Chandy, R., Dorotic, M., Krafft, M., & Singh, S. S. (2010). Consumercocreationinnewproductdevelopment.JournalofServiceResearch,13(3),283–296.

Juster, F. T. (1969). Consumer anticipations and models of durable goods demand. InMincer,J. (1969).EconomicForecastsandExpectations.NewYork:NationalBureauofEconomicResearch.

Kumar,V.,Bhaskaran,V.,Mirchandani,R.,Shah,M. (2013).Creatingameasurable socialmediamarketingstrategyforHokeyPokey:IncreasingthevalueandROIofintangibles&tangibles.MarketingScience,32(2),194–212.

Kumar,V.,Donkers,A.C.,Venkatesan,R.,Wiesel,T.,&Tillmanns,S.(2010).Undervaluedor overvalued customers: Capturing total customer engagement value. Journal ofServiceResearch,1(3),297–310.

L’Hoest-Snoeck,S.,VanNierop,E.,&Verhoef,P.C. (2015).Customervaluemodelling inthe energy market and a practical application for marketing decision making.InternationalJournalofElectronicCustomerRelationshipManagement,9(1),1–32.

Leeflang, P. S. H., Verhoef, P. C., Dahlström, P., & Freundt, T. (2014). Challenges andsolutionsformarketinginadigitalera.EuropeanManagementJournal,32(1),1–12.

Leenheer,J.,Bijmolt,T.H.A.,VanHeerde,H.J.,&Smidts,A.(2007).Doloyaltyprogramsreallyenhancebehavioral loyalty?Anempiricalanalysisaccounting forself-selectingmembers.InternationalJournalofResearchinMarketing,24(1),31–47.

Lewis,M. (2006).Customer acquisition promotions and customer asset value. Journal ofMarketingResearch,4(2),195–203.

Li, H. A., & Kannan, P. K. (2014). Attributing conversions in a multichannel onlinemarketing environment: An empirical model and a field experiment. Journal ofMarketingResearch,5(1),40–56.

Rapp,A.,Baker,T.L.,Bachrach,D.G.,Ogilvie,J.,&Beitelspacher,L.S. (2015).Perceivedcustomershowroomingbehaviorandtheeffectonretailsalespersonself-efficacyandperformance.JournalofRetailing,91(2),358–369.

Risselada,H.,Verhoef,P.C.,&Bijmolt,T.H.A.(2015).Indicatorsofopinionleadershipincustomernetworks:Selfreportsanddegreecentrality.MarketingLetters.Forthcoming.

Rust,R.T.,Lemon,K.N.,&Zeithaml,V.A.(2004).Returnonmarketing:Usingcustomerequitytofocusmarketingstrategy.JournalofMarketing,68(1),109–127.

Rust,R.T.,Zeithaml,V.A.,&Lemon,K.N.(2000).Drivingcustomerequity:Howcustomerlifetimevalueisreshapingcorporatestrategy.NewYork:TheFreePress.

Shah,D.,Kumar,V.,Qu,Y.,&Chen,S. (2012).Unprofitablecross-buying:Evidence fromconsumerandbusinessmarkets.JournalofMarketing,76(3),78–95.

Sharp,B.(2010).Howbrandsgrow.Australia&NewZealand:OxfordUniversityPress.

Simon,C.J.,&Sullivan,M.W.(1993).Themeasurementanddeterminantsofbrandequity:Afinancialapproach.MarketingScience,12(1),28–52.

Sriram, S., Balachander, S.,&Kalwani,M.U. (2007).Monitoring the dynamics of brandequityusingstore-leveldata.JournalofMarketing,71(2),61–78.

Trampe,D.,Konuş,U.,&Verhoef,P.C. (2014).Customer responses toannelmigrationstrategiestowardthee-channel.JournalofInteractiveMarketing,28(4),257–270.

VanDoorn, J.,Lemon,K.E.,Mittal,V.,Naβ,S.,Pick,D.,Pirner,P.,Verhoef,P.C. (2010).Customer engagement behavior: Theoretical foundations and research directions.JournalofServiceResearch,13(3),253–266.

Venkatesan, R., & Kumar, V. (2004). A customer lifetime value framework for customerselectionandresourceallocationstrategy.JournalofMarketing,68(4),106–215.

Venkatesan,R.,FarrisP.,&WilcoxR.T.(2014).Cutting-edgemarketinganalytics,realworldcases and data sets for hands on learning. Upper Saddle River,New Jersey: PearsonEducation.

Verhoef,P.C.(2003).Understandingtheeffectofcustomerrelationshipmanagementeffortson customer retention and customer sharedevelopment.Journal ofMarketing, 67(4),30–45.

Verhoef, P.C.,Kannan, P.K.,& Inman, J. (2015). Frommulti-channel retailing to omni-channelretailing: Introductiontothespecial issueonmulti-channelretailing.JournalofRetailing,9(2),174–181.

Verhoef,P.C.,Neslin,S.A.,&Vroomen,B.(2007).Browsingversusbuying:Determinantsof customer search and buy decisions in amulti-channel environment. InternationalJournalofResearchinMarketing,24(2),129–148.

3Data,dataeverywhere

Introduction

In thisworldofbigdataeverythingstartswithdealingwith theoverloadandvarietyofdatathatareavailable,tohelprealizetheambitiousobjectivesofvaluecreationwithbigdataanalytics.Andthat’swhydataisatthecoreofthisbook.Earlier,weexplainedwhywetalkabout“big”data(byusingthethreeVsmodel).Oneofthethree‘V’swasvariety.Thisrepresentsthefactthatdatanowadaysarearisingfrommoreandmoresources,andthatinfactdatacanbefoundeverywhere.Inthischapterwewillelaborateonthisvarietyofdatasources.Wewilldiscusswhatkindofdatasourcescanbedistinguishedandhowallthesedatacanbecategorized,aswellashowdatashouldbeprocessedandstored.Specificattentionwillbegiventodataqualityanddatasecurity.Wewillfollowthesamestructureasinotherchapters,meaningthatwewilldiscussmarketdatasources,productandbranddatasources,anddatasourceswithcustomerdata—allwithaspecificfocusonnewdatasourcesthathavebecomeavailable.

InChapter3.1wewilldiscussdataintegration:howtoextract,transform,andloaddata,how to create all types of variables in the commercial data environment and how tophysically integrate data sources from different aggregation levels. In Chapter 3.2 thesubjectwill be big data privacy issues and security. For companies, aswell as for theircustomers,privacyandsecurityareveryrealthemesthatarebeingdiscussedalmostdailyonTVand innewspapers—anddefinitely shouldbediscussedwithin the context of thisbook.

Datasourcesanddatatypes

Firmshavemanydifferentdatasourcesattheirdisposalforfillingtheircommercialdataenvironment. By a commercial data environment we mean the technical infrastructurewherealldata is stored,processed,andaccessed, coming fromallkindsof sources tobeused for commercial steering. In this environment, we see nowadays a combination of“traditional” datawarehousing technology combinedwithmodern big data tooling, likeHadoop, MapReduce, etc. This combination of technology enables dealing with thechallengesposedbytheexistenceofdifferenttypesofdatasources,andspecificchallengesindatavolumeandvariety.InChapter5wewilldiscussthismoreindepth.Withinthesedatasources,wedistinguishbetweeninternalversusexternaldatasourcesandstructuredversus unstructured data sources. The different combinations of internal versus externalandstructuredversusunstructureddataareshowninFigure3.1.

Figure3.1Twodimensionsofdata:Datasourceversusdatatype

Source:AdaptedfromNairandRaman(2012)

Afirstimportantdistinctioninthesedatasourcesisbetweenexternalandinternaldatasources.

Externaldatasourcesversusinternaldatasources

External data sources are not presentwithin the firm and are frequently purchased andadded to the commercialdata environment from (or collectedby) externaldatavendors.ImportantexamplesofthesedataareZIP-codeorhouseholddatasets.InZIP-codedatasets,informationisprovidedonthecharacteristicsofhouseholdslivinginthatZIP-code,suchasaverageincomelevel,educationlevel,averagehouse-price,etc.Thisinformationcanbelinked to individual customers through the ZIP-code of the customer. More and morevendors are transforming their ZIP-code datasets into household datasets, by enrichingZIP-code datasets with information obtained from omnibus questionnaires and publicly

availabledatasetsforcommercialuse(suchasinformationonthesaleofhouses,includinghouse prices). This transformation helps to createmore powerful profiles of households,insteadofhavingtomakeuseofextrapolationsforallhouseholdsinaZIP-code.

Otherexternalinformationmay,forexample,involvethefinancialcreditworthinessofacustomer. Credit card firms such asAmerican Express,Mastercard, andVisa have goodknowledge of whether a consumer is creditworthy, by analyzing the use of their creditcard.Moreandmorefirms(e.g.Mastercard)arestartingtoexploittheirdatasetstobeusedas an external data source for checks of creditworthiness, for example with telecomproviders.

Important suppliers of external consumer data are Experian, Axciom, and Nielsen-Claritas.ExternalB2BdataareprovidedbyfirmssuchasGraydonandDun&Bradstreet.Externalconsumerdataarefrequentlyusedfortargetmarketingpurposesandforprofilingcurrentcustomers in termsof socio-demographics, lifestyle,andpsychographicvariables.SupplierssuchasNielsen-Claritashavedevelopeddetailedsegmentationsystems,inwhicheach household belongs to a certain psychographic-lifestyle segment, such as youngfamilies,ruralfamilies,youngurbanprofessionals,etc.(seeFigure3.2).

Anotherkindofexternaldataismarketingresearchdata,mostlycollectedforfirmsbytheir market research agencies. These agencies often have large panel sets, withrespondents in place to perform quantitative market research on topics such as brandperformance, campaign evaluation,media effectiveness, and product innovation.Marketresearchcanalsoberealizedbyusingdataofthefirm’sowncustomers,suchascustomersatisfactionorNPS,usuallycollectedbyexternalmarketingresearchsuppliers.Dependingonprivacyregulations,customerpermission,and/orthespecificresearchprofessionalcode,marketresearchdatacanbeincludedinthecommercialdataenvironmentaswell.

Furthermore, firms may collect competitive intelligence data on competitive actions.These data are usually not internally present, and can be collected by marketingintelligencedepartments.Researchhasshownthatcompetitiveactions,suchascompetitiveadvertising,haveanimpactoncustomerbehavior(Prins&Verhoef,2007).

Figure3.2ExampleofNielsen-ClaritasinformationforaNewYorkZIP-code

Source:Claritas.com1

Anothertypeoffast-growingexternaldataaredatafromsocialmedia.ThesedatacomefrompartieslikeFacebook,Twitter,LinkedIn,andInstagram,allofwhichhavehugeuserbases,sothatoftenasubstantialproportionofafirm’scustomersparticipateinthesesocialmedia.Linkingthesedatatothecommercialdataenvironmentisquiteachallengeandnotyetacommonpractice.Thisismainlyduetothehighlyunstructuredwaythesedataarebeing created. Firms are collecting these data by using platforms like Radian6 for socialmedia monitoring, but are still struggling to interpret the data and integrate theseplatformswiththeircommercialenvironment.

Internaldatasourcesarealreadypresentwithinthefirmandmayincludepoint-of-salesdata, transaction data, invoice data, contact data, and usage data. These internal datasourcesaregatheredandstoredinthedatawarehouse.Internaldatacanbeconsideredtobe very powerful. The possession of information from internal data is important ifcustomer behavior is to be described, understood, and predicted. However, these datasources need a lot of data preparation in order to create valuable information. Withincustomer value management (CVM) this is the most frequently stored data (Verhoef,Spring,Hoekstra,&Leeflang, 2002).Alsomanymodelsaiming tooptimize thecustomervalueonlyuseinternaldata(e.g.Venkatesan&Kumar,2004).

Structuredversusunstructureddata

Analternativeway tostructuredatasources is tomakeadistinctionbetweenstructuredandunstructureddatasources.Structureddataaredatathatcomeinafixedformat,basedonadetailed record andvariable structure, good labelingof values in thedatabase, andhigh data quality. The invoice datamentioned earlier (internal data) and ZIP-code data

(externaldata)aregoodexamplesofhighlystructureddata.Ontheothersideofthecoinwehaveunstructureddata.Thesedataareoftenverybulkyinsize,withoutafixedformat,containingalotoffreeformattextandoftenneedagooddealofdatainterpretationanddata reduction inorder to createusable information.Examplesof thesedata sources aredata from customer contact (internal data), where customers and their questions orremarks are often registered in free format text, and social media data (external data)contained inTwittermessages,Facebookcomments, etc.Figure3.3 shows an example ofhowunstructureddatacanbetransformedintostructureddata.

Figure3.3Illustrationofstructuredandunstructureddata

Aspecialremarkshouldbemadeaboutmobiledata.Mobiledata(inFigure3.1depictedasinternal,unstructureddata)isarathernewphenomenonthattookoffwiththeincreaseof smartphone and tablet penetration. This kind of data is quite unique since it offersorganizationsthepossibilitynotonlytomonitorwhatcustomersaredoing,butalsowheretheyaredoingit—therebygivingbigopportunitiestoallkindsoflocation-basedservices.Installed apps give the customer access to products and services that offer all kinds ofserviceandtransactionfunctionalities.Sincetheusageofdatageneratedbytheseappsisoftennotyetintegratedwithinthearchitectureofthecommercialdataenvironment,andbecausethedatafromtheseappsareoftenunstructured,thereisstillpioneeringworktobedoneinthisarea.

Marketdata

Anotherlensfordiscussingdatasourcesisourbreakdownbymarket,brand/product,andcustomer.Thisisnotonlyusefulfromtheperspectiveofthebigdatachallengebutalsoitcreates awareness about the differences between the data sources for market,product/brand,andcustomers,differencesthatariseduetoscope,detail,andpowerofthedatasourceathand.Wedefinetwotypesofmarketdata:

1. Marketdataonthesupplyside:Datathatdescribeandexplainmetricslikemarketsize,marketvolume,marketshare,marketdevelopment,mediaspendetc. forallplayers inthemarket. Ideallywiththepossibilitytobeshownpercompetitor in

themarket.2. Marketdataonthedemandside:Datathatdescribetheconsumersinthemarket,

theirbuyingandspendingbehavior,theirsocio-demographics(e.g.age,householdsize,income),andtheirneeds.

Marketdataonthesupplysideisoftencollectedbyagenciesthataggregate,forexample,salesfiguresfromdifferentsuppliersinordertocreateamarketoverview.Sometimesthiscanbeanagencythatoperatesforthewholeindustry.Forexample,employmentagenciesintheNetherlandsalldelivertheirplacementsper4-weekperiodtoABU,anorganizationthatcollectsthisdataforthewholeindustryanddeliversthisbackatanaggregatedlevel.AnotherexampleisNielsen,whichcollectsscanningdataandcreatesdetailedsalesfiguresbasedonthis(seeFigure3.4).

Marketdataonthedemandsideismainlycollectedbyresearchagencies.AnexampleisGfK,whichcreates,forinstance,anoverviewoftheinsurancemarketbyaskingconsumersdetailsoftheirinsurancepapers.Themaindifferencewithsupply-sidedataishowthedataare collected. Demand-side data are built up from the users or buyers of products orservices,oftenatthehouseholdlevel,asfarasitconcernsconsumers.Wecanconsidertheexternal data as provided by vendors like Experian (see above) also as an example ofmarket data, since these are built up from the consumer/demand-side perspective at thehouseholdorZIP-codelevel.Marketdataonthedemandsidearenotonlydataonproductusageorproduct-buyingandsocio-demographics,theycanalsoconsistofmore“softdata,”describing consumer needs and values as they exist in the market. These datasetscategorize consumers in segments like “cosmopolitans” or “post-materialists” as used by,for example,Motivaction, aDutch research agency. These groups display specific needswithrespect toacertain industryorproductcategory. InFigure3.5 thementalitymilieusegmentationofMotivationisdisplayedasanexampleofmarketdataonthedemandside.

Figure3.4ExampleofmarketdataonthesupplysideforUKsupermarkets

Source:adaptedfromNielsenTotalTill,NielsenHomescan(2014)2

Figure3.5Exampleofmarketdataonthedemandside

Source:Motivaction3

Bydefinitionmarketdatahavethebroadestscope,astheiraimistorepresentthetotalmarketforanindustryorproduct/service.However,sincecollectingthiskindofdatacanbeverycostly,thedatamaysometimesnotbeasdetailedasdesiredandsometimesalsobeoutdated,sinceregularlyupdatingiscostlyaswell.Thisofcourselimitsthepoweroftheuseofthesekindsofdatasources.

Bigdatainfluenceonmarketdata

Alternatives for traditional market data are arising, especially due to developments likeopen data projects, where governments and organizations share their data using theenormouspowerofdigitalizationandonlinedata.Asanexample,byanalyzing thedatafrom the app called “Slice” (an app that helps customers to keep track of purchases andstore receipts) a very good estimate wasmade of recently sold units of the newApplewatch.Anotherexampleisdatacollectedbyallkindofcomparisonsites.Theyallhaveagoodperspectiveonmarketvolumesandtransactions,especiallysincetheonlineshareoftransactionsinallindustriesisstillrisingveryfast(mainlyintheorientationphaseofthebuyingprocess)andpossiblebiasesarereducing.

Branddata

Forbranddataweagainmakeadistinctionbetween:

1. Branddataonthesupplyside:Datathatdescribethevolumeandmarketshareofspecificbrands/products

2. Branddataon thedemandside:Data thatdescribeandexplainshow(potential)

customersjudgeacertainbrand.

Branddatahavealimitedscopeinthattheyfocusonlyonbrandsandtheirperformance.Collectingthesekindsofdata,especiallyforthedemandside,isacostlyeffort.Tocollectthese data, research is necessary, and this brings up discussions about the quality andvalidityofresults.

Forbranddataon the supply side thereare actuallyno specificdata sourcesbuiltup.Either these data are a subset or cross-section of the above-mentionedmarket data (forexample,tocalculatethemarketshareofaspecificbrand)ortheyareextractedfromthesystemsthatstorethesalesperbrandfortheorganization(seediscussionofcustomerdatalater in thischapter). InFigure3.6we showanexampleofwhat this typeofdatamightlooklike.

Branddataonthedemandside(seetheexampleinFigure3.7)focusesonthedifferentstepsinthecustomerorientationprocess,asmeasuredbybrandfunnelswithstepssuchasbrandawareness,brandconsideration,andbrandpreference(seeChapter2.1).Thesekindsof data are collected by conducting (online) market research that is performed inlongitudinalstudies.

Figure3.6Illustrationofbrandsupplydataextractedfrominternalsystems

Figure3.7Illustrationofbranddemandbasedonmarketresearch

Source:adaptedfromUBSEvidenceLab4

Bigdatainfluenceonbranddata

Inthesedaysofbigdata,branddatacanalsobecollectedbylisteningintosocialmediatomeasure brand sentiment, not to replace but to add to branddata collectedby research.Reviewsandratingscanalsobeanimportantbigdatasourceformeasuringandcollectingbranddata.Topreventbiaseswesuggestthatthesekindsofnewdatasourcesareanalyzedincombinationwithtraditionalsources,insteadoffullyreplacingthem.

Customerdata

Althoughthedistinctionbetweensupplyanddemanddatacanalsobemadeforcustomerdata,thereisasmuchcustomerdataavailableonthesupplysideasthereisonthedemandside:

1. Customerdataonthesupplyside:Datathatdescribesthe(historical)productandservicesusedbythecustomerduringtherelationshipwiththeorganization

2. Customer data on the demand side: Data that describes the expectations,satisfaction,andinteractionsofthecustomerwiththeorganization.

Customerdataisdifferentfrommarketandbranddatainseveralways:customerdataisoften very detailed and accurate (especially since billing data is the life-blood of theorganization)andbydefinitiononly covers customers (and sometimesprospectsandex-customers).So ithasanarrower scopeandhasbecome (fromthemomentorganizationswere successful in extracting and storing this data) very powerful in the commercialsteeringprocess.

Customer data on the supply side are mainly stored in the CRM systems of theorganization (see Figure 3.8 for an example of a relational marketing databaseenvironment). In a contractual setting the billing data are an important source foridentifying the customer and assessing the financial value of the customer, as well ashelpingtoanalyzeproductholdingandusage.Inanon-contractualsettingwherecustomeridentification is not always possible these data are often limited to data on the productlevel,statingusageandbuyingofproducts,includingrepeatpurchase.

Customerdatafromthedemandside(seeFigure3.9)comefrommarketresearchonthecustomer base, doing research onmetrics such asNPS or customer satisfaction, or frominteractionsbetweenthecustomerandtheorganization,forexampleinacallcenter,shop,

oronawebsite.Becausethesedataareoften(at leastpartly)unstructured, interpretationandanalysisofthemisnecessarytotransformthemintoinformationandknowledge.

Bigdatainfluenceoncustomerdata

Theinfluenceofbigdataoncustomerdatacanbefoundinthelargeinternal,unstructureddatasetsthatarebeingbuiltupandthataremoreandmorewithinthescopeofthedataanalyst. Also, the online presence of organizations designed to help them to serve theircustomers, is something that is creatingbig data—for example, in the “my [organizationname]”environmentthatmanyorganizationscreatetoinformtheircustomersabouttheirbilling, their products, etc., not to mention all kinds of “self-service” options. This isinfluencingcustomerdataonthesupplysideaswellascustomerdataonthedemandside.

Figure3.8Illustrationofadatamodelofcustomersupplydata

Usingthedifferentdatasourcesintheeraofbigdata

Toassessandtomakemaximumuseoftheaddedvalueofeverydatasource,weusethe5“W”s(seeFigure3.10):

Whoisthecustomer?Whatisthecustomerdoing/using?Whereisthecustomerusing(orbuying)theproduct?Whenistheproductboughtorused?Whyistheproductused?

Fromacommercialperspective,thesearethequestionsthatalwayspopup.Itshowsthateverydatasourcehasitsspecificstrengthinansweringcertain“W”questions,depictedbythe colour in Figure 3.10. So survey data, for example, is very strong in answering the“Why”ofcustomerbehavioraswellasin“What”thecustomerisdoingorintendingtodo.However, these “W” questions are often only answered from a single data sourceperspective.Thiscancauseproblemsbecausetheanswerstothedifferent“W”smightnotbe consistent, or theymay even be contradictory. To solve this, we think that the datasources thatmight come from surveys, transactions, socialmedia, ormobilesneed tobecombinedwitheachother.

Figure3.9Illustrationofcustomerdemanddata(NPS)

Butcombining thesedatasources isquitechallenging froma technical, statistical,andlegal perspective. From a technical perspective, because some data sources might beanonymousordonothaveauniquekeytolinkthesources.Fromastatisticalperspective,because not all data sources have the same coverage of the population (this might, forexample,beapopulationsampleoronlydataoncustomersandnotonthetotalmarket).To deal with this, extrapolation techniques will be necessary. From a legal perspective,because combining sources on the individual level might create conflicts with privacy

policies,professionalcodes(suchasthatofmarketresearchers),orlegislationonprivacy.InChapter3.1wewillgo intomoredepthonhowtosolvemergingdatasets fromdifferentorigins,withdifferentcoverageand/orlegalhurdlesformerging,usingtechniquessuchasdatafusion.

Figure3.10The5“W”smodelforassessmentofdatasources

Datawarehouse

Thedatawarehouseisoftenconsideredtobethecentralelementofthecommercialdataenvironment. Data sources from inside and outside the firm provide data to the datawarehouse.Throughadatagatheringsystem,thedataaretransformedintoanelectronicmedium.Thedatawarehousefunctionsto:

1. preparethedataforstorage2. storethedata3. literallydescribethedata4. manageandcontrolthedata.

From the data warehouse, analytical databases or data marts are provided to customerintelligencespecialists,datascientists,anddatabaseanalysts(Zikmund,McLeod,&Gilbert,2003). The data warehouse can be enrichedwith a software-based information deliverysystem. This information delivery system transforms data into information. Thisinformationismainlydescriptiveinnature,andmayincludemetrics,suchasthenumberof active customers, the churn percentage in the last month, or the average sales percustomer. However, the information delivery module may be enriched with standardmodules for more advanced analytical purposes. Most frequently used are campaignmanagementtools,inwhichstandardizedandoftenautomatedanalyticaltoolsareusedtoselectcustomersformarketingcampaigns.

Databasestructures

Withinafirm,severaldatabaseswithrelevantcommercialdataareusuallypresent.Firmsmay have databases on point-of-sales information (e.g. store sales; sales through salespersons), products, invoicing, customer contacts, etc. In a datawarehouse all these datasourcesareintegratedintoonelargecustomerdatabase.Thecentralfocusofthisdatabaseis the customer. However, from this database should not only customer information beretrieved, but also other information, for example information on sales personperformance.

Animportantelementofgooddatabasesisthatthedataarearrangedinsuchawaythattheycaneasilyberetrievedbyusers.Thedatabasestructuremay,however,dependontheuser. A salesmanager is probablymost interested in the performance of the sales reps,whileaproductmanagermost likelywants toknowtheperformanceof theproducts.Acustomer manager is most likely interested in customer metrics, such as customerprofitability.Toovercomethishurdle,relationaldatabasestructuresarecurrentlystandard.Relationaldatabasesusedifferentkey-variableswhichlinkseveraldatabasestoeachother.Forexample, inan insurancecontext thecustomer idandtheproduct idareusuallykeyvariables.InaB2Bcontextthecustomeridcanagainbeakeyvariable,whilethesalesrepid can be too. Based on the type of information required, a primary key variable, asecondarykeyvariable, etc. canbedistinguished. Ina salesdatabase, the salesperson idwouldbe theprimarykeyvariable,while in a customerdatabase the customer id is theprimarykeyvariable.

Figure3.11showsanexampleof thestructureofacustomerdatabase. In thisexamplethecustomer id is theprimarykeyvariable.ForCVMpurposes thecustomerasprimarykeyvariableisofessentialimportance.Thecustomernameprovidesfurtherinformationonthatcustomer.Thetypeofproductisasecondarykeyvariable,andthetransactionchannelis another. One can easily add more key variables, such as membership of a loyaltyprogram, time, etc. From this database, one can in principle derive a databasewith theothernon-primarykeyvariablesaskeyvariables.Forexample, ifonewantedtotakethetype of product as a key variable, a database could be created through aggregation andtransformation procedures which shows per product the number of customers and themostfrequentlyusedtransactionchannels(seeFigure3.12)

Havingmultiplekeyvariables,thedatabasecanbetreatedinamulti-dimensionalway.For example, one canhave twodimensions per customer, by knowingwhichproduct inwhichyear(time)theypurchased.Workingwithmoredimensionsquicklybecomesmorecomplicated. Customer information users generally find analyses at a high dimensionalleveldifficulttounderstand,althoughsoftwarevendorshavedevelopedmulti-dimensionaldatabasesinanefforttoovercomethis(Zikmund,McLeod,&Gilbert,2003).

Figure3.11Exampleofsimpledatatablewithcustomerascentralelement

Figure3.12Exampleofproductdatatablederivedfromcustomerdatabase

Dataquality

Dataqualityconsistsofseveraldimensions:

CompletenessofdataDatabeinguptodateNomistakesindata

Completenessofdatareferstowhetherallavailabledataarepresentforallcustomers.Forexample, adataacquisitionchannelmaybepresent foronlya sampleof customers (e.g.Verhoef&Donkers, 2005).Mistakes frequentlyoccur, especiallywith regard to customerdescriptors,suchasnameandaddress.Thesemistakesmayariseascustomerswritedownunclearnamesetc.onforms,perhapsonpurpose,orwhentypographicalerrorsmeanthatdataentryisnotdonecorrectly.

The data being up to date concerns whether the data are being updated on a veryfrequentbasis.Adatabasethatisnotuptodatecaneasilycontainmistakesonallkindsofvariables. For example, if customers havemoved to another address, the address in thedatabasewillbewrongifithasnotbeenupdated.Orifintegrationofdatabasesisnotdonefrequently,arecentproductpurchaseorarecentdefectionmightnothavebeenincludedyet,leadingtounreliableinformation.

Mistakescanpotentiallyhavestrongnegativereputationalconsequencesforafirm.Forexample, a firmmay continue sendingmail to someonewho has recently passed away.Moreover, data being not up to date may also cause wrong predictions to be made onfuturecustomervalue,whichmightleadtolessthanoptimalstrategies.

Data quality is thus important for shaping performance in CVM strategies (Zablah,Bellenger, & Johnston, 2004). However, the question is whether firms should have anextremely high level of data qualitywith perfectly complete data, no datamistakes and100%up-to-datedata.Neslinetal. (2006)proposethattheremightanoptimumlevel (seeFigure3.13).Achievinggooddataqualitycomesata cost.However, thesecosts rise inanon-linear fashionwhen higher levels of data quality are achieved, while the return intermsofincreasedcustomervaluefromthisdataqualitymayactuallydecreaseinanon-linearfashion.Thisimpliesthatfirmsshouldassesstheoptimumlevelofdataqualityandnotpursuea100%score.

Thereareseveraloptionsavailabletosolveproblemswithdataquality.Oneoptionistousereferencedatabasesforcleaninghistoricaldataqualityissuesinthedatabases;theycanalso be used as a reference table during data entry.Another option is to use a softwareapproach for data cleaning, for example, byusing tools to recognize duplicateswithin adataset,ortorecognizedifferentoptionsforregisteringaspecificaddress.

Figure3.13Netbenefitsofinvestingindataquality

Source:AdaptedfromNeslinetal.(2006)

Missingvaluesanddatafusion

Unfortunately,datawillneverbeperfect.Oneoftheimportantproblemsresearchersmayface is that therearemissingvalues inthedata.Forexample, there isno informationongenderandage for somecustomers.Oneeasyway todealwith thesemissingdata is tothrowawayobservationswithmissingvalues.Thismay,however,causesampleproblems,especiallywhenthesemissingvaluesoccurveryfrequentlyinanon-randomfashion.Forexample,mainlyforcustomersinaNortherngeographicarea,socio-demographicdatamaybemissingbecausecustomerswerenotasked toprovide thesedata.Missingvaluesmayalsohaveameaning.Forexample, customersmayonpurposenotprovidedata,becausethey distrust an organization or they are not very loyal customers (Vroomen, Donkers,Verhoef,&Franses,2005).Inthelattercasemissingvaluescanpotentiallybepredictorsofbehavior. For example, onemight assume that customers withmissing values aremorelikelytodefect,astheydistrustthefirm.Thusingeneraloneshouldcarefullyanalyzethereasonformultiplemissingvalues,andconsiderwhetherthesemissingvaluesoccurinarandom or non-random fashion. If it is a random event, one could probably deleteobservationswithmissingvaluesorreplacethemissingvalueswith,forexample,themeanvalue.

Arelatedproblemisthatmanyfirmshavespecificdataforonlyasubsetofthesample.Forexample,supposesatisfactiondataareonlypresentfor10%ofthecustomerdatabase.For the other 90% no satisfaction data are available. Still, one might like to know thesatisfactionscoresforthesecustomersaswell.Inthesameveinonecouldhaveshareaofwalletdataforasampleofthecustomerbase.Fortunately,methodshavebeendevelopedbywhichtheinformationinthesamplecanbeusedtoestimatethemissingdataintherestof the customer base. These techniques are called data fusion techniques (Kamakura &Wedel,1997).ArathersimpleformofthesetechniquesisreportedinDonkersandVerhoef(2001). Theyuse a regressionmodel to predict the total number of insurance products acustomerpurchases.This includespurchasesatboth the focal firmandoutside it—i.e. atcompetitors. For more information on advanced data fusion techniques we refer toKamakuraandWedel(2003)andKamakuraetal.(2005).

Conclusions

In this chapter we started with making the distinction between different data sources,grouping them in different ways: structured versus unstructured and internal versusexternal.Westressedthatbigdataarenotjustabouttheexternalunstructuredsources,butalsohavetodealwiththeothersources. In linewiththeotherchapters inthisbook,wedistinguishedthreetypesofdata:marketdata,brand/productdata,andcustomerdata.Foreverytypethereisademandsidetypeofdataandassupplysidetypeofdata.Further,wediscussedtheaddedvalueofeverysourceofdata,byanalyzingthedifferent“W”s(Who,What,When,Why,Where).Thedescriptionofdatabasestructuresandpossibleissueswithdata quality andmissing values should help in tackling practical issues aroundworkingwithdata.

Notes

1 See www.claritas.com/MyBestSegments/Default.jsp?

ID=20&menuOption=ziplookup&pageName=ZIP%2BCode%2BLookup#(accessedSeptember18,2015).

2Seehttp://meteorpublicrelations.com/wp-content/uploads/2014/09/Supermarket-sales-growths-show-improvement-for-

first-time-this-year.pdf(accessedSeptember18,2015).

3 For more information on mentality groups: www.motivaction.nl/en/mentality/the-mentality-groups/ (accessed

September18,2015).

4 See http://appleinsider.com/articles/14/12/01/ubs-survey-finds-10-of-consumers-want-a-smartwatch-expects-24m-

apple-watch-sales-in-fiscal-2015(accessedSeptember18,2015).

References

Donkers,B.,&Verhoef,P.C.(2001).Predictingcustomerpotentialvalue:Anapplicationintheinsuranceindustry.DecisionSupportSystems,32(2),189–199.

Kamakura,W.A.,&WedelM.(1997).Statisticaldatafusionforcross-tabulation.JournalofMarketingResearch,3(4),485–498.

Kamakura, W. A., & Wedel, M. (2003). List augmentation with model based multipleimputation:Acasestudyusingamixed-outcomefactormodel.StatisticaNeerlandica,57(1),46–57.

Kamakura,W.,Mela,C.F.,Ansari,A.,Bodapati,A.,Fader,P.,Iyengar,R.,Naik,P.,Neslin,S. A., Sun, B., Verhoef, P. C., Wedel, M., & Wilcox, R. (2005). Choice models andcustomerrelationshipmanagement.MarketingLetters,16(3/4),279–291.

Nair,R.,&Narayayan,A.(2012).GettingResultsfromBigData:ACapabilities-DrivenApproachtotheStrategicUseofUnstructuredInformation,Booz&Hamilton.www.strategyand.pwc.com/media/file/Strategyand_Getting-Results-from-Big-Data.pdf(accessedNovember26,2015).

Neslin,S.A.,Grewal,D.,Leghorn,R.,Shankar,V.,Teeling,M.L.,Thomas,J.S.,&Verhoef,P. C. (2006). Challenges and opportunities in multichannel customer management.JournalofServiceResearch,9(2),95–112.

Prins,R.,&Verhoef,P.C.(2007).Marketingcommunicationdriversofadoptiontimingofanewe-serviceamongexistingcustomers.JournalofMarketing,71(2),169–183.

Venkatesan, R., & Kumar, V. (2004). A customer lifetime value framework for customerselectionandresourceallocationstrategy.JournalofMarketing,68(4),106–215.

Verhoef,P.C.,&Donkers,B.(2005).Theeffectofacquisitionchannelsoncustomerloyaltyandcross-buying.JournalofInteractiveMarketing,19(2),31–43.

Verhoef,P.C.,&Langerak,F. (2002).Elevenmisconceptionsabout customer relationshipmanagement.BusinessStrategyReview,13(4),70–76.

Verhoef,P.C.,Spring,P.N.,Hoekstra,J.C.,&Leeflang,P.S.H.(2002).Thecommercialuseof segmentation and predictive modelling techniques for database marketing in theNetherlands.DecisionSupportSystem,34(4),471–481.

Vroomen, B., Donkers, B., Verhoef, P. C., & Franses, P. H. (2005). Selecting profitablecustomersforcomplexservicesontheInternet.JournalofServiceResearch,8(1),37–47.

Zablah, A. R., Bellenger, D. N., & Johnston, W. J. (2004). An evaluation of divergentperspectivesoncustomerrelationshipmanagement:Towardsacommonunderstandingofanemergingphenomenon.IndustrialMarketingManagement,33(6),475–489.

Zikmund, W. G., McLeod, Jr. R., & Gilbert, F. W. (2003). Customer relationshipmanagement:Integratingmarketingstrategyandinformationtechnology.NewJersey:Wiley&Sons.

3.1Dataintegration

Introduction

Although thebiggest challenges in thebigdata era seem to lie in collecting and storingdata,webelievetherealchallengeistheintegrationofallkindsofdatasourcestorealizesuccessful big data value creation. This is because many of these data sources are notmeantorbuiltupforthepurposeofintegrationwithotherdatasources,andalsobecausethe data sources (integrated or not) often contain data variables that need furtherprocessingtocreateuseful information.Inthischapterwewilldiscussthedifferentstepsthat are necessary for data integration as well as the process of creating marketingvariablesoutofthedifferentavailabledataitemswithinthecommercialdataenvironment.

Integratingdatasources

The first step in the process of data integration is also calledETL (see Figure 3.1.1), theprocess of extraction, transformation and loading the input data sources into the datawarehouse.ThisprocessofETLcaneitherbedonebyhand,orbyusingdifferenttypesofprogramming(e.g.SQL),orbyusingoff-the-shelfsoftwaretools(e.g.Informatica).

Figure3.1.1TheETLprocess

Extraction

Thecrucialpartof theextractionstage is theselectionof therelevantdatasources tobeincludedinthedataenvironment.Partofthisprocessisvalidation:checkingwhetherthedatasourcesextractedaretherightonesthatareneededandwerespecifiedinitially.Thisvalidation is necessary since time changesmight occur in the input of data sources thatmightaffectthefurtherprocessingoftheinputdataandeventhepossiblerelevanceofthespecificdata source.Anotheraspectof theextraction step ishow thedataareextracted.Thiscanbedoneviaanautomatedquerythatrunsataspecifiedfixedmomentintimeorit can be performed manually (not preferred because of possible human errors beingintroduced,andthetimeit takes).However, inthefirstphaseofsettingupanintegrateddataenvironmentandbuildingupasortofa“proof-of-concept”wherethedatastillneedtobeextractedorarenotyetfullystable,itmaybepreferabletoperformseveraliterationsby hand. The frequency of extraction of new datasets is also something that should beconsideredwhen organizing the loading process. Typically, formarketing information amonthly refreshment should be suitable, although especially in highly dynamicenvironments a higher frequency should be considered. Real-time refreshment does notseem feasible, from a technical perspective (there will always be a delay from dataprocessing) and also from a business perspective. Analysts should ask the question ofwhether the extrabenefit gained fromhavingmoreup-to-datedata isworth the cost ofcollection (e.g. Neslin et al., 2006). Those insights that are needed in real time can betriggeredinanoperationalenvironmentbyimplementingbusinessrulesand/oralgorithmstomakemaximumuseofcollectedinsights.

Transformation

After extracting the data, the next step is transforming the different data sources to fitthemintothecommercialdataenvironment,byapplyingallkindsofbusinesslogics.Thetransformationofthedataduringthetransformationphaseisoftendesignedtomakethedata easier to handle (more storage efficient and more robust in data quality). Typicaltransformationsareselectingcertaindataattributes,replacingvaluesinthedata(e.g.“M”instead of “Male”), translating values from character into numeric (e.g. “1” instead of“activecustomer”),calculatingvalues(revenueexcludingVAT=revenuedividedby1.21),or extracting and splitting variables into different variables (e.g. taking out the housenumber and street nameof the address field andputting it into twonewvariables).Allstepsinthetransformationstagearedesignedtomakethedataeasiertohandleduringtheanalysis, make storage easier, and make integration with other data sources morestraightforward(bymakingthedatasourcesuniform).

Load

Aftertransformingthedatatheywillbefurtherprocessedbyloadingthemintothedesireddata environment. Thiswill normally be donewith an intermediate step called staging.During staging all data sources are loaded in an intermediate environment before beingloadedintothetablesofthedesireddataenvironment.Thebenefitofthisisthatitmakesitpossible to do final checks before publishing the data into the environmentwhere usersactuallystartworkingwiththedataorwherereportsaregenerated.Thestagingareaalsoservesassortofabackupincasesomethinggoeswrongbyloadinginthefinaltables.Mostof the time thedata in the stagingarea is intermediateandwillbeerasedwhen it isnolongerneeded.There are twoways bywhich thedata canbe loaded into the final dataenvironment. The first is by overwriting the earlier dataset in the commercial dataenvironment.However, this couldmean losingall kindsofhistoricaldataand isnot thepreferredoption.Anotherwayistoappendthenewdataduringtheloadingprocesstotheexistingdatainthecommercialdataenvironment.Inthiswayhistoricaldatacanbekeptandnewdatapointsareaddedeachtimethedataarerefreshed.

Dealingwithdifferentdatatypes

Datasourcesvaryintermsofcontent,scaling,source,andpresencewithinthecommercialdataenvironment.Itisourstrongbeliefthatinsteadoffocusingonthe“V”(volume)ofbigdata, the real challenge is addressing the “V” (variety) of data sources, especially aswebelievethateverysourceaddsaspecificdimensionto thecommercialdataenvironment.Webroadlydistinguishbetweenfourdatatypes(seealsoFigure3.1.2):

Declareddata(customerdescriptors)Appended data (transaction and billing data, customer contact data, marketingcontactdata,customercharacteristics,customerattitudes,brandperformancedata)Overlaiddata(zip-code,householddataandresearchdatalikebrandperformance,customerattitudes,marketandcompetitordata)Implied data (segmentation, scoring models, share of wallet, recency frequencymonetaryvalue(RFM)classificationetc.).

Declareddata:Customerdescriptors

Customerdescriptorsinvolveallcharacteristicsrequiredforcontactingcustomers,suchasname, address, zip-code, phone-number, and email address. Using this informationmarketing actions can be directed at individual customers and invoices can be sent tocustomers.

Appendeddata

Figure3.1.2Thedifferentdatatypes

Transactiondataconcernalldataoncustomers’(financial)transactionswiththefirmandmayinvolvevariablessuchasthelasttimeofpurchase,thetypeofproductpurchased,themonetary value of the purchase, transaction channel, and product returns. Customercontact data (contact history) are customer-initiated and may, for example, involveinformation requests, complaints, clarifications on invoices, website visits, and contactchannel (i.e. phone, email etc.). Marketing contacts are firm-initiated contacts and mayinvolve the number ofmailings sent, timing ofmailings, and details of loyalty programmembership. Customer characteristics concern additional information on the customerrelated to socio-demographics, psychographics, lifestyle etc. Part of this data may arisefrom internal sources.Forexample, forhealth insurancecustomershave toprovide theirsex,age,householdsizeetc.

Overlaiddata

External data suppliers also provide information, at different levels of aggregation. Onespecific external profiling analysis is the incorporation of Zip-codes. External dataproviders,suchasAxciomandExperian,havespecificZip-codelevel(orevenhousehold)information. Using this information firms can gain an understanding about which Zip-codes(andthuslocal/regionalareas)over-orunderrepresenttheircustomers.Alongwiththis zip-code/household information, these external data suppliers have also developedinformation on specific segments, such as “rural families” and “single households.” InFigure3.1.3weprovideanoverviewofthesegmentsthatareusedbyExperianUK.Firmscanusethisinformationtofurtherexternallyprofiletheircustomersandcustomergroups.Forexample,anonlineretailermayfindthat“ruralfamilies”areover-representedintheircustomerbase,whilethe“singlehousehold”isalmostcompletelyabsent.Indexingthedata(furtherdiscussedinChapter4.1)canbeveryuseful.

Figure3.1.3OverviewofsegmentationschemeusedbyExperianUK

InFigure3.1.4weshowanactualexampleofaclothingretailerthathasmanystoresinlarger villages outside the city. This retailer aims to compare his clientele with thepopulation.Ascanbeobserved fromtheanalysis, the“prestigepositions,”and“transient

renters” have the highest over-representation in the customer base. The stores do notattractsegmentssuchas“modesttraditions.”

Figure3.1.4ExternalprofilingusingZIP-codesegmentationforclothingretailer

Source:AdaptedfromExperian,UK1

Customer attitude data are data on customer satisfaction and other attitudes, such ascommitmentornetpromoterscore(NPS).Thesedataareusuallycollectedbycarryingoutsurveysamongsamplesofthecustomerbase.Asaconsequence,thesedataareusuallynotpresent for all customers. Brand performance data (i.e. data on brand awareness, brandpreferenceetc.)aremeasuredforcustomersandnon-customersandareonlyavailableforasubset or sample of the customer base.Market and competitor data are data onmarketshare, market volume, volume share, as well as information on the performance andmarketpositionofcompetitors.

Implieddata

Finally, firms can derive data from combining all these other data. Important derivedvariables are share ofwallet, propensity to buy scores, credit scoring, churn probability,andcustomerlifetimevalue(CLV).Infactderiveddatacanalsobeconsideredasimplicitinformation. The created variables are based on calculations, assumptions, andcombinations of data sources. Here, the creativity and capabilities of data-scientists arecrucialincreatingacompetitiveedgeontheuseofdata.

Figure3.1.5PresenceofdatatypesforDutchfirms

Sources:Verhoefetal.(2002)andVerhoefetal.(2009)

Verhoef,Spring,Hoekstra,&Leeflang(2002)andareplicationstudy(Verhoef,Hoekstra,& Van der Scheer, 2009) have investigated the presence of data types for Dutch firms.Customerdescriptorsandtransactiondataarethetypesofdatamostfrequentlystoredincustomerdatabases.Between2003and2008astrongincreaseinthepresenceofalltypesofvariablesisobserved(seeFigure3.1.5).Thisreflectsthefactthatfirmshavebeensuccessfulinsettingupcompletedatawarehouses,inwhichinformationforallkindsofdatabasescannowbeintegrated.

An important element of the commercial data environment is the customer database.Nowadays,firmshaveverylargedatabasesofcustomerdataattheirdisposal.Thesedataarisefromallkindofsources.Inthepastfirmsusuallyhadseveraldatabaseswithdifferentkindsofcontents,whichwereusedinbusinessprocessessuchassendingoutinvoicesandhandling of inbound phone calls. Moreover, usually the data were structured aroundproducts instead of around customers. These databases were not integrated. Instead,individualcustomerdatawerefragmentedovermanydatabases.Asaconsequencefirmsdidnothavedetailedindividualcustomerinformationandcouldnothavethefullpictureofwhatwasactuallyhappeningwitheachcustomerovertime.Duetothe integrationofdatabases in many large-scale CRM projects, one integrated customer database is nowfrequentlyavailable.

Dataintegrationintheeraofbigdata

Thestepsdescribedabovearetypicalforaconventionaldatawarehouseinacommercialsetting. The data in a typical data warehouse often have a customer/prospect-centricapproach,where all data are being collectedwith thepurposeof linking it to individualcustomers or to prospects at the individual or household level. Each table in the datawarehousecontainsakeydata field thatultimately links themto thecustomer level.Bycombiningallthisdatafromthedatawarehouse—ideallyintoonesingletableorflatfile—with still rather raw data added into the data warehouse, the dataset becomes richer,especiallyif(asdescribedabove)newvariablesarecreated.Thenextstepistodealwithallkindofdatasourcesthatalsocanbestoredinthecommercialdataenvironmentbuteitherhaveinformationofonlyasubsetofcustomersorwerecollectedwithanotheraggregationlevel inmind (aswas the case for themarket andproduct datadescribed inChapter3).However,thetechnicalchallengeofdealingwiththedifferentaggregationlevelsisnottheonlychallengeformakinggooduseofthesecombinedsourcesthatarenotintegrated.Inthissection,aswellasdiscussinghowthesedatasourceswithdifferentaggregationlevelscanbetechnicallyintegrated,wewillalsodiscussotherchallenges.Weintendtolookatanewwayofanalyzingthesesourceswithaviewtoworkingouthowtointegratethem.

Non-integrated customer, market, and product data are typically used by differentdepartmentsintheorganizationmakinganalysesandreportsonthesedata,everyonefromtheperspectiveoftheirown“silo.”Thisoftenresultsinalotofconfusionandannoyancewiththeend-usersoftheoutputs.End-usersmakecommentslike:“Wehavetonsofdata,sowhyisittakingsomuchtimetocreatetherightinsightswhenweneedthem?”or“Whydowehavetogatherourcrucialmarketinginsightsfromsomanydifferentdepartmentswithin the organization?” or “We are overloaded with reports and overviews, but theydon’t give us input on how to improve our business performance” or “Although I nowunderstandwhathashappened,pleasetellmealsohowtoact.”

Underlyingthesecommentsthereareinfactthreechallengesorganizationshavetodealwithinordertomakemaximumuseofintegrateddata(seeFigure3.1.6):

Technicalchallenges:ScatteredandfragmenteddatasetsAnalytical challenges: Lack of a similar and synchronized customer/consumerperspectiveandconsistent segmentation;explaining thepast insteadofpredictingthefutureand;howtoanalyzewithdatasetsatdifferentaggregationlevelsBusiness challenges: No link with metrics of the P&L to realize alignment andacceptancefromthefinancialdepartment.

Asimplifiedexamplemightillustratethisbest.Let’sassumeweareaninsurancecompanyselling car insurance. The management needs insights that help to explain the currentperformance of the organization and its competitive strength, including the levers for

performanceimprovementandactionableinitiatives.Thecustomerdataisavailableattheindividual customer level, with a customer-id and name and address of the customerindicatingtheproductandbrandthecustomerbought,whenitwasbought,whereitwasbought and at what price, and how the customer is using the product. For this sameexamplewealsohavemarketdataavailableperbrandforallmarketplayerspermomentintime,indicatingtheaveragevalueperbrandpercustomerandthenumberofcustomersper brand. Themarket data have been collected bymarket research through a randomsample of the population of consumers with car insurance. The third relevant dataenvironmentisonlinedatastoredonthewebserver.

Figure3.1.6Thechallengesofdataintegration

Thetechnicalchallengesofintegrateddata

Thetechnicalchallengeofdataintegrationismainlycausedbydifferentaggregationlevelsofdata sources, but alsoby the fact that, as inour example above, the customerdata islikelytobestored intheCRMdatawarehouse,accessiblebyanalysts fromthecustomerintelligence department, while the market data is stored by the research agency (onlydistributingareporttoitsclient)orwithinthemarketresearchdepartment.Soinfactthereare two technical challenges. The first challenge is to realize a centralized storageenvironment for the different data sources from different departments, the integratedenvironment thatwe call the commercial data environment. The second challenge is tophysicallyintegratethedatasourcestoenableanalysisacrossallthesedatasources.

Thefirsttechnicalchallengecaneasilybesolvedbyspecifyingthedesireddataextractsand downloading them into a dedicated storage environment. This can be a dedicatedserverorastand-alonepieceofhardware,likeapowerfulpc.Especiallyinthephasewhereproofisneededforabigdataapproach,thelastoptioncanbeanaffordablealternativetoafullyfledgednewcommercialdataenvironment.

Thesecondchallenge is thephysical integrationof thedata sources.Thereare severalwaysoftacklingthis,eachwithitsadvantagesanddisadvantages.Thedifferentoptionsare

mainlydrivenby theaggregation levelof thedatasource: thedatasourceat thehighestaggregation level can be considered “the weakest link,” because it determines theaggregation levelallotherdatasourceshavetoalignwith. Inourexampleabovefor theinsurancecompany,itisverylikelythatthemarketdatatypicallyanalyzedatthemarketplayerlevelisthehighestaggregateddatasource.

Integrationattheindividuallevel

Thefirstoptionfortechnicalintegrationofthedatasourcesiswhatwecall“integrationattheindividuallevel.”Thismeansthattheindividual(orread“household”for“individual”)isidentifiedineverydatasourceandthatwestartlinkingthedifferentdatasourcesusingthekeysidentifyingtheindividual:thekeysforcombiningtwosourcesshouldcorrespond.Integratingthesourceswillusuallystartwiththelargestdatabaseattheindividual level.Thefirststepisthusintegratingthisdatabase(oftenthecustomerdatabase)withthenextlargestdatabase.Let’sassumeinourexamplethatwewanttointegratethecustomerdatawiththeonlinedata.Inthecustomerdatatheindividualcanbeidentifiedbyacustomerid,orforexampleacombinationofaddressplusbirthdata.Thenextstepistoidentifyinouronlinedata the rightkey for integrating individualonlineusers to the customerbase. Inouronlinedatawewilloftenfindthedatastoredatthevisitlevel.Sothefirstthingtodoisto identify unique visitors. This should result in a unique visitor id resulting from acombinationofacookieid,andanipnumber(andforregisteredusersthatarecustomersthiscouldbethecustomerid).Theregisteredusersshouldbeintegratedwiththecustomerdatabasebyusingthecustomerid.Thenon-matchingcustomersshouldbestoredwiththecustomeridastheuniqueidentifierandthenon-matchingwebsitevisitorsshouldbestoredwiththeuniquevisitorid.Sinceintegrationattheindividuallevelinvolvesprivacyissuesand sensitive information, one should consider specific measures (for examplepseudonymizing,asdescribedinChapter3.2).

Integrationattheintermediatelevel

The secondoption, integrationat an intermediateaggregation level,usesa segmentationbased on dimensions that can be identified in all sources to be integrated. Thesegmentation then becomes the common denominator to which all data should beaggregated for every time period. In our example of the insurance company, we coulddefine thedimensionsasageand income.Classifyingeachof these twodimensions in5classeswouldresultin25segmentstobeidentifiedineverydatasourcepertimeperiod.Ifadatasourcedoesnothavethisinformation,thisproblemcouldbesolvedbyfirstaddingexternaldatabasedonZip-codeorhouseholdlevel(e.g.fromsomewherelikeExperian).

Integrationatthetimelevel

The third option for data integration is the least advanced. In this option data will beaggregatedtothetimeperiodthatcanbeidentifiedinthedatasourcesandthetimeaxiswillbethedimensiononwhichtocomparethedifferentdatasources.

Mixtureofintegrationoptions

Thefourthoptionisperhapstheonethatwillbeusedmost—acombinationofoptions1,2and3.Inthisoptioneverydatasourcetobeintegratedwillfirstbecheckedforavailablepossibleuniqueidentifiers.Basedonthisassessmentitwillbecomeclearwhichwillbethebestoftheabovethreeoptionstouseperdatasource.

Theanalyticalchallengesofintegrateddata

Aswellasthetechnicalchallengestobeaddressed,wealsoseefouranalyticalchallenges.Thefirst is thesynchronizationofthecustomerandconsumerperspectivestobeusedintheintegrateddatasetandinthesegmentationtobeusedacrossandwithinthedifferentdatasets. Defining the customer level and the consumer level can be difficult. Oneimportant decision to be made is whether to define this at the household level or theindividual level. Once that is decided the next problem is how to identify this levelconsistentlyacross thedifferentdatasources.Thisasks for internalalignmentwithintheorganization and departments and clear business logics to apply the final choices.Assumingthistobethecase,thenextanalyticalchallengetobedealtwithisthechoiceofthedimensionsandcalculationofthesegmentationtobeused.Thesourcetobeintegratedcouldalreadyhaveasegmentationappliedbywhichresultsarepresented,andtheymaybevery specifically tailored to the information in that specific database. When usingsegmentationintheintegrateddata,eitherfortheintegrationitselforforanalyzingacrosssegmentsacross thedata sources, auniformwayof segmenting thedata isneeded.Thisrequires a generic approach to the dimensions of segmentation appliedwithin a specificdata source. Just as with the decision onwhat customer/consumer definition should beused,segmentationisalsoatopicthatshouldbealigned,discussed,andappliedintherightwaywithintheorganization.Thethirdanalyticalchallengearoundintegrateddataisthetypeof insights tobegenerated.Since integrateddata, combiningcustomer,market, andbrand input, shouldofferanextrabenefitover reportingon thedifferentdatasources inisolation, we think that the ultimate goal should be not trying to explain what hashappened. The real benefit should be in being able to make a prediction on what willhappen,bycreatingaholisticviewofthemarketingperformanceandidentifyingtheleversthatexplainhistorical,current,andfutureperformance.Thefourthanalyticalchallengelies

in the way integrated data can be used for modeling purposes, especially since in theaggregations necessary for physical data integration part of themodeling power of datamightbelost.WewillgointomoredepthonthisinChapter4.2,whenwediscussspecificbigdataanalytics.

Thebusinesschallengesofintegrateddata

Weseetwodifferentchallengesaroundintegrateddatafromthebusinessperspective.Animportant business challenge is to define the right key performance indicators (KPIs)within and across the different data sources. The KPIs to be defined should reflect themarketing performance around customers, brand, and market, and should point upopportunities for improvement. The next challenge is to link theKPIs to the company’sP&L.

Conclusions

Inthischapterwehavediscussedthreesubjectsarounddataintegration:theETLprocess,the process of creating all kinds of new variables from combining datasets, and thedifferent challenges around the integration of data sources with different aggregationlevels.Althoughallthreesubjectsarecrucialformakingbigdataintegrationasuccess,weconsider the last subject (and then especially the technical challenge of dealing withdifferent aggregation levels) as driven by all developments around big data. Solving thechallengesarounddifferentaggregationlevelsisthewaytodealwiththe‘V’ofvarietyinbigdataandsoisakeysuccessfactorinrealizingthepotentialofbigdata.InChapter6wewill discuss the case of an insurance company to give a pragmatic example of how thechallengesarounddataintegrationcanbedealtwith.

Note

1Seewww.experian.co.uk/marketing-services/products/mosaic-uk.html(accessedSeptember20,2015).

References

Neslin,S.A.,Grewal,D.,Leghorn,R.,Shankar,V.,Teeling,M.L.,Thomas,J.S.,&Verhoef,P. C. (2006). Challenges and opportunities in multichannel customer management.JournalofServiceResearch,9(2),95–112.

Verhoef, P. C., Hoekstra, J. C., & Van der Scheer, H. R. (2009).Competing on analytics:StatusquovancustomerintelligenceinNederland.ReportofCustomerInsightsCenter(RUGCIC-2009-02),UniversityofGroningen.

Verhoef,P.C.,Spring,P.N.,Hoekstra,J.C.,&Leeflang,P.S.H.(2002).Thecommercialuseof segmentation and predictive modelling techniques for database marketing in theNetherlands.DecisionSupportSystem,34(4),471–481.

3.2Customerprivacyanddatasecurity

Introduction

Sofar inthisbookwehavemainlydiscussedthevalueopportunitiesofbigdataandbigdataanalytics.Indeed,thesevalueopportunitiescanbeconsiderable.However,inaneraofbig data firms are confronted with concerns about the storage and usage of data. Thisspecifically concerns customer data. If customers have used digital and mobile devices,theywillprobablyneverbeanonymousagain.Theirbehaviorislikelytobetracedonline,butalsooffline.Forexample,ifcustomerswithamobiledeviceenterastore,retailersusingWiFi-basedtrackingtoolscanfollowcustomersinthestoreandhowtheyshop.Onamoreglobal level there are continuous debates about customers’ data being analyzed bygovernments. This has become high profile news as a result of documents leaked byEdwardSnowden,whichrevealednumerousglobalsurveillanceprograms,manyofthemapparently run by the US National Security Agency (NSA) with the cooperation oftelecommunicationcompaniesandEuropeangovernments.Thishasraisedahighlevelofconcerngloballyon the informationprivacyof individual global citizens.Can firms likeGoogle,Facebook,Microsoft,andAmazonbetrustedwithregardtotheirprivacypolicies?And is the information contained in emails sent inGmail,MicrosoftOutlook, or Yahooobservableandavailableforanalysisbygovernments?

Thebigdatamovementhasthuscreatedastrongerdebateonprivacy.ThecounselloroftheUSPresident,JohnPodesta,suggeststhatbigdataraisesseriousquestionsonhowweprotectourprivacy.TheGermanPrimeMinisterAngelaMerkelalso considersprivacyabigissue,butbelievesthatdespitetherisingprivacyconcern,wemustbeabletousebigdata toour advantage.Unfortunately, this isnot so straightforward.Recently, theDutchBank INGRetail announceda test that involved sharingpaymentdatawithother firms,with the objective of improving the value delivered to customers through providingattractive and relevant offers to ING customers. This resulted in strong reactions fromcustomers, stakeholders, and politicians. Even the president of theDutchNational Bankactively communicated that he was not in favor of this big data initiative. ING had toretract this initiative and reduce their big data ambitions substantially. This case onlyemphasizesthestrugglethatfirmsarehavingwiththeprivacyaspectofbigdata,althoughtheprevalenceofprivacy issuesmayvarybetween firmsand sectors.However, it isnotonly privacy that is an issue. Data security also deserves close attention, as data couldbecomeaccessibletootherpartieswithmischievousorcriminalmotivesthroughhackingofcomputersystemsandunsecureddatatransport.

Inthischapterwewillmainlyfocusonprivacy.Wewilldiscusswhatprivacyactuallyisandhowcustomersconsiderprivacyissues.Wewillalsodiscusstheroleofgovernmentsandspecifically legislation.Bigdataprivacypolicieswillbeelaboratedon.Datasecuritywillbeconsideredinthelastsectionsofthischapter.

Whyisprivacyabigissue?1

As discussed above, big data has stirred up the privacy discussion. Privacy has been anissuefordecades,butthebigdatadevelopmenthasputprivacybackontheagendaoftopmanagementandgovernments.AccordingtoJones(2003)bigdataareconsideredtobeathreattoprivacyforseveralreasons:

BigdataarepermanentlyavailableBigdatainvolvelargevolumesofdataBigdataoncustomersarecollectedinvisiblyThereisnogoodassessmentoftheprivacysensitivityofdataDueto the largevolumeofdata, it isno longeraccessibleandunderstandablebycustomersDatafrommultiplesourcesarebeingfusedCustomersperceivealackofcontrolonbigdatacollection.

If theprivacydebate,particularly inWestern societies, receivesgrowingattentionand ifmorerestrictivepoliciesaredeveloped,theconsequencesforbigdataanalyticsandvaluecreationarelikelytobeconsiderable.TheBostonConsultingGroup(2012)2hascalculatedthat the so-called digital identity of customers has a value of approximately 330 billioneuros for firms,while forcustomers thevalue isdoubled,withavalueofapproximately670billioneuros.Theyalsoestimatethatalackoftrustandprivacycandestroy440billioneuros in thevalueofcustomers.Ofcourseonecandebateabout theexact figuresof theestimatedvalueconsequences.However,thesecalculationsclearlyshowthatthereismuchatstakeforbothfirmsandcustomers.Firmscanbecomemorerestrictedinwhattheycandowithbigdata,whichmay,forexample,resultinlessefficientandeffectiveadvertising.Customers can be worse off, because they get less attractive personalized offers andpersonalized services may be reduced. Despite this, governments are rightfully worriedaboutprivacy issuesand theuseofdata.Thekey issue isprobably that there is societaltrade-off between what we deem as important data- and privacy requirements and theassociatedbenefitsfromalackofprivacy.

Whatisprivacy?

Wetalkalotaboutprivacy,butwhatisitactually?Privacyisaconceptthathasastrongphilosophical background. In essence, it implies the right to be left alone (Warren &Brandeis,1890).Thisalmostsuggeststhefactthatoneshouldbeabletoliveanonymouslywithoutbeingdisturbedbyanyonefromanyinstitution.Thisisofcourseratherunrealisticfor thevastmajorityof consumers.However, it clearly suggests the fact that individualsshouldbeabletomakethetrade-offbetweenseclusionandinteraction(Westin,1967).Thisdiscussionisstillratherphilosophical.BringingitmoreintothecontextofdataGoodwin(1991)suggests thatanindividualshouldbeconsciousaboutwhichdataandinformationare being shared and towhat extent they control this sharing. Two concepts are rathercrucialhere:consciousnessandcontrol.

A concept that is frequently discussed is “privacy concern.” Privacy concern usuallyfocusesonsixaspectsofdata:

1. Datacollection:“Toomuchdataandinformationisbeingcollected”2. Datausage:“Dataisbeingusedforotherpurposesthanservingtheconsumer”3. Datamistakes:“Mistakesindatacanhavenegativeconsequences”4. Datainfringement:“Unauthorizedaccessandusageofdata”5. Datacontrol:“Insufficientcontroloverowndata”6. Dataconsciousness:“Notsufficientlyinformedondatapolicies.”

Research on privacy has mainly considered privacy concern as the main topic ofinvestigationandconsidereddriversofprivacyconcernand theconsequencesofprivacyconcerns.Wewillelaborateonthisinthenextsection.

Customersandprivacy

Customers can thusbe concernedabout theprivacy consequencesof bigdata.Concernsandfearsaremainlyrelatedtothere-usageofdata,reputationlossesduetotheuseofdata,and the wrong interpretation of data and information. Research also suggests thatcustomers are unconscious or not aware of the usage and collection of information andtheyactuallyseemtobeworriedthattheydonothavesufficientknowledgeondatausage(Hong&Thong,2013).However,marketresearchhasalsoshownthat therearedifferentsegmentsofcustomerswithdifferentviewsonprivacyconcerns (MarketResponse,2014;Fletcher,2003;Ackerman,Cranor,&Reagle,1999):

Fundamentalist: Fundamentally against the collection andusage of data by firmsandgovernmentsPragmatist:WillingtoshareinformationaslongasbenefitsarereceivedUnconcerned:Donotconsidersharinginformationasaproblemandconsideritaspartofdailylife.

RecentresearchintheNetherlandssuggeststhatthemajorityofcustomersarepragmatist,althoughasubstantialminorityarefundamentalist.Unconcernedcustomersrepresentthesmallest group (Market Response, 2014). Substantial research has devoted attention tospecificdriversofprivacyconcerns.Thesestudiessuggestthat, forexample,olderpeopleare less likely to share information,and femalesaremoreconcernedabout theirprivacy(Youn,2009).

An important question is whether privacy concern also leads to less sharing ofinformation by customers. The relationship between privacy concerns and actual datasharing behavior is not so obvious, and the correlation between privacy concern andbehavior is low.Onlywhencustomershavevery strongconcernsabout theirprivacydotheychangetheirbehavior(VanDoorn,Verhoef,&Bijmolt,2007).Youcouldactuallyarguethatcustomersdonotbehaveveryconsistently.Customersseemtobeworriedabouttheirprivacy, but billions of customers around the globe constantly share very personalinformation on Facebook, and leave digital traces behind online. The strong disconnectbetweenprivacyconcernandactualbehaviorisreferredtoasthe“privacyparadox.”

Governmentsandprivacylegislation

Governments across the globe develop their own legislation. There is a good deal offragmentation on privacy regulation. The most severe laws are found in the EU andCanada.TheUSAandAustraliahavelesssevereprivacylegislation.Inemergingcountries(i.e.Brazil,Russia, India,andChina,knownasBRIC) legislation is relatively limited (seeFigure3.2.1).

Figure3.2.1Dataprotectionlawsaroundtheglobe

Source:AdaptedfromDLAPiper,Dataprotectionlawsoftheworld3

Within the European Union, the membership states have developed Data ProjectionDirectives, leading to EU general data protection regulation. There is an official EUinstitution that oversees the implementation of this legislation by firms that has theauthority to provide officialwarnings, audits, and finally fineswith amaximum of 100millioneurosor5%of thecompany’sworldwideturnover. Indeed, theEUhasstartedupprivacy trajectorieswith importantplayers, suchasGoogle.Therearesomeprinciples intheEUlegislationthatshouldbeemphasized:

1. Right to be forgotten: Customers have the right that their information can bedeletediftheyrequireit

2. Right of portability: Customers have the right to view their collected personalinformation

3. Informed consent: Privacy is default for customers.Data are only collected if acustomeroptsforthat(opt-in)

4. Clearnotice:Thereisanofficialsavetimeofcontactinformation5. Privacy by design: Protection of privacy should be part of thewhole design or

engineeringprocessofnewproductandserviceinnovations.

Asnoted,thelegislationintheUSAisnotasstrongasitisintheEU.Oneofthelargestdifferences lie in the approach.Whereas theEU is rather proactive in its legislation, the

USAisratherpassive.Beyondthat,thelegislationintheUSAdiffersbetweenstates.IntheUSAtheFederalTradeCommittee(FTC)isresponsibleforgeneralprivacylaws.TheFTCwrote a large privacy report in 2012. They concluded that industry efforts to addressprivacythroughself-regulation“havebeentooslow,anduptonowhavefailedtoprovideadequate and meaningful protection.” The report recommended many actions. Some ofthem would bring US regulation more in line with that of the EU. For example, theyrecommend a privacy by design approach. Furthermore, they emphasize that customersshouldhaveachoice.Customersshouldbepresentedwithchoicesaboutthecollectionandsharingoftheirdataatthetimeandinthecontextinwhichtheyaremakingdecisions—notafterhavingtoreadlong,complicateddisclosuresthattheyoftencannotfind.FTCalsorecommends a “do not track” mechanism, which governs the collection of informationabout consumers’ Internet activity to deliver targeted advertisements and for otherpurposes.4Sofar,thereporthasonlyresultedinrecommendationsforfirms,whicharenotyetimplementedinUSAlegislation.

Beyonddatalegislation,firmsshouldbeawarethattheresultingpoliciesfromtheirdatausagearealsoaffectedbylaws.Forexample,iffirmstargetpersonalcharacteristics,suchasreligion,genderandrace,thiscouldresultindiscrimination.Althoughthefirmcouldactin accordancewith specific privacy legislation, its resulting policies could be in conflictwith regulations on discrimination. Hence firms should look beyond the privacylegislation,andshouldalsoconsiderregulationrelatedtospecificmarketingactions.

Privacyandethics

Onahigherlevelonecoulddebatewhetherfirmsshouldmainlyfocusonlegislationandtakethatastheruleonhowfartheywillgowithdatacollectionorwhethertheyshouldtake a broader perspective. The latter leads to a discussion on how firmsmake “moral”decisions.Onecouldadvocateanapproach inwhich firmsareallowed tocollectdataaslong as the law allows them to do so. However, one could also adopt a more ethicalperspective that goes beyond laws, where firms consider that they have an ethicalresponsibility. De Bruin (2015) distinguishes two important dimensions in this respect:ethical decisionmaking andmoral intensity. Ethical decisionmaking applied to privacyinvolvesfourstages:

1. Firmshavetorecognizethatprivacydecisionshaveamoraldimension2. Firmsthenhavetoformanethical judgmentconcerningwhatoughttobedone

withregardtoprivacy3. Firmshave toestablish themoral intention toact inconformitywithwhat they

havejudgedtobetherighttypeofbehaviorwithregardtoprivacyissues4. Firmshavetoengageinthatbehavior.

The “moral intensity” of an ethical issue such as privacy refers to themagnitude of theconsequences of the actions and the probability of it arising, as well as whether theconsequencesareconcentratedonagroupofpeopleordispersedamongthem.Importantly,it also depends on whether there is any social consensus about the fact that particularactions are good or evil (DeBruin 2015).5DeBruin (2015) explicitly states that, roughlyspeaking,whenevilconsequencesarelikelyorsevere,affectpeopleincloseproximityoralargenumberofpeople,andwhenthefirmperceivesthistobethecase,theissue’smoralintensity ishigh.Themoral intensityofan issuedetermineshow firmsproceed througheachofthefourethicaldecision-makingstages.Ifmoralintensityishigh,thenanissueisconsideredasamoralissueandelaborativeethicaldecisionmakingisrequired.

Theabovediscussionisrathertheoretical.Takingaprivacyperspective,inpracticefirmsshouldputconsiderationofthemoralintensityofprivacyissueshighontheiragenda.Itisour contention that the big data development has created a wider acknowledgement ofprivacyissuesinsociety.Itisprobablytoomuchtosaythatbigdatamayresultinwrongor evil consequences. However, when data are accessible to criminals “evil” thingsmayactually happen to customers. Overall, we believe that privacy is an ethical issue thatrequiresmoreattention thanonlyconsidering the law.Thiswouldalsobe in linewithamore customer-centric approach, as firms caring for customers and the interest ofcustomersshouldstronglyconsidertheirviewsonprivacyandhowtheyshoulddealwithdata. Moreover, recent discussions have shown that big data initiatives can potentiallyharm a firm’s reputation. In sum, privacy issues surrounding big data should be a very

important discussion issue within firms, and they should go through more intensivedecisionmakingthanonlyconsideringavailablelegislation.

So in general we recommend that firms should strongly consider the ethical andreputationalconsequencesof theirbigdataandprivacypolicies.Specifically, theyshouldadopt stakeholder management and consider reactions from customers, the government(includingpolitics),andthemedia.

Privacypolicies

Therehasbeenextensive researchonprivacyand specificallyprivacypolicies.Basedonthisresearchwealsohavesomerecommendationsonhowtodealwithspecificdataandprivacyissues:

Onlycollectdata thatare relevantandcongruent.Relevantdatameansdata thatare considered useful in servicing the customer. Congruency of datameans datashould be related to the product or service provided. For example, for financialservices, data on ownership of insurance products or financial transactions arecongruent with the service. However, data on medical issues would not beconsideredcongruent.Fromaprivacyperspectivetherule“moreis less”holds.Themoreinformationisaskedfor,thelesscustomersprovide!Sofirmsshouldbelimitthemselvesinwhattheyaskcustomerstoprovide.Give something back to customers. Data has value for firms as well as forcustomers. Rewards (monetary and non-monetary) can inhibit the customer’swillingness to share information. However, this does not work for irrelevantinformation.Moreover,therearedifferencesbetweencustomersegments.Be transparent on data usage. Providing a clear privacy statement positivelyinfluencesdatasharing.Itmayreducecustomers’lackofawarenessondatausage,whichmayreduceprivacyconcern.Communicate the specific benefits of sharing data. If customers perceive thebenefitsof sharingdata, theyaremore likely to share.Privacymainlybecomesaproblemwhentheadvantagesofsharingdataarenotclear.Invest in brand trust.Customers aremore likely to share informationwith firmstheytrust,astheybelievetherisksinvolvedarelower.Trustparticularlybecomesan issue when customers are asked to share personal and sensitive information.Even trusted firms thatwronglyuse the provideddata can find that their use ofdatabackfiresonthem.Onefinalrecommendationis thatfirmsshouldgivecustomerscontrol!Accordingto the Boston Consulting Group, 82% of customers want to have control of thecollectionandusageoftheirinformation.Customerswhofeelincontrolhavemorepositiveviewsonsharinginformationandsharemoreinformationaswell.Thiscanstrongly increase the effectiveness of marketing. Giving control to customersimplies the use of opt-in and opt-out options and the use of permission-basedmarketing.

The power of giving control to customers is excellently shown in a study byCatherineTucker(2014)ofMIT.ShereportsresultsofastudyamongFacebookusers,whereasimplecontrolbuttononprivacywasadded.Privacybecamestandardandfriendswerenolonger

visibletoeveryone.Moreover,opting-outfortheuseofdatawasmademoreconvenient.Tucker (2014) reports that the effectiveness of targeted and personalized advertising inparticularwasmorethandoubled(seeFigure3.2.2).

Privacyandinternaldataanalytics

Privacy also has consequences for how data is analyzed, as the analysis of data on anindividual customer level might not be allowed due to legislation. We consider thefollowingspecificsolutions:(seeFigure3.2.3)

Ifindividualdataarereallyproblematictoanalyzefromaprivacyperspective,onecould aim to analyze data on higher aggregation levels. For example, specificmarketsegmentscouldbestudiedoraggregatedanalysescouldbedone.Alltheseanalysescanprovidecustomerinsights.

Figure3.2.2EffectivenessincreaseofFacebookadvertisingcampaignsafteradditionofprivacybutton

Source:AdaptedfromTucker(2014)

When analyzing data, individual data can be anonymized. This is typically donewhen one only requires customer insights. Anonymizing of data is usually thestandard in traditionalmarket research. If one aims to include the results of theanalysisinthedatabase,themodelresults(e.g.churnmodel)canbeusedtocreatethemodeloutcomeinthedata(e.g.churnprobability).Aspecific formofanonymizing thedata iswhatwecallpseudomyzing thedata.Withaspecifickeydataareanonymized.Thisanonymizationisdonebyatrustedparty (e.g.external ITfirm,external law-firm).Onlythis trustedpartyknowsthekey.Theanalystsexecutetheanalysesandtheresultscanbeinputintothenormalcustomerdatabase.Again,theexternaltrustedpartytakescareofde-anonymizingofthedata.A final technique is linked to the already discussed permission-basedmarketingapproach (Godin, 1999). A subsample of not anonymous data on customerswhohavegivenpermissiontoanalyzeandusetheirdataisanalyzed.Forthecustomersgrantingpermission,theresultsoftheanalysescanbeincludedinthedatabase.Forthe remaining part of the data the outcomes of the analysis can be included byusingthemodelestimationstopredictthespecificvalues.Oneconcernhereisthat

thepermission-basedsamplewilltypicallynotbearandomsampleandhencetheanalyticalresultscouldbebiasedduetosomeself-selectionissues.

Onefinalconcernisthat,althoughwenowadayshaveconsiderablediscussiononbigdatadevelopments, European firms in particular hesitate to collect longhistories of customerdata. For example, telecom firms inWestern countries keep historical data for only onemonth and then throw the data away. This limits the possibilities of carrying outmoredynamicanalysesthatinvolvepaneldata(Prins&Verhoef,2007;Ascarza&Hardie,2013).Somenewmethods are beingdeveloped that aim to benefit fromhistorical data, but byusing earlier analytical results. Specifically, Holtrop, Weiringa, Gijsenberg and Verhoef(2014)havedevelopedtheso-calledGMOK(generalizedmixtureofKalmanfilters)model.Thisisstillinitsinfancyhoweverandrequiresadvancedmethodologies.

Figure3.2.3Differentwaysofhandlingprivacysensitivedata

Datasecurity

A subject closely related to privacy is data security. The risk of data infringement (theunauthorized access and usage of data) discussed in the previous section can seriouslyimpactonanindividual’sfinancialand/orpersonalsafety,andalsoonthecontinuityofthefirmwitha securitybreach.Remember, for example, thehackingofLinkedInpasswordssomeyearsago,ortheAppleiCloudhackwherenudepicturesofcelebritieswerespreadallaroundtheInternet.Theseexamplesshowthatevenrenownedmajorcompaniescanhavedata security problems, and highlight the need for every firmworkingwith privacy orcompanysensitiveinformationtotaketherightmeasurestosecuretheirdata.

Wedistinguishbetweenthreeelementsofdatasecurity.Thefirstisthepeopleelement,peoplethatuseorhaveaccesstothedata.Thesecondisthesystemelement,dealingwiththephysicaldatastorageandtheenvironmentwherethedataarestored.Thethirdistheprocesses, meaning the procedures and policies (including penalties) as defined by theorganization, that define the rules for access, continuity, steering, and monitoring ofsecurityperformance.

People

Notsurprisingly,peopleareoftentheweakestlinkindatasecurity.Veryoftenemployeesofthefirmare(atleastpartly)responsibleforsecurityissues.EverybodyknowsexamplesofsensitivedataonaUSBstickbeingleftinapublicplaceoronanon-companycomputer,oremailattachmentsbeingsentoutside theorganization to thewrongperson.Theseareonlyexamplesofdatasecurityfailuresthatarenotintentional.Evenbiggerarethesecuritythreats from people both inside and outside the firm who have malicious or criminalintentions. This means that firms should protect themselves from human failure ormisbehavior.Theyshouldscreentheiremployees’credentialsandmakeemployeesawareofthecompliancepoliciesinplace—evenletthem(someofthematleast)takeanexamonthecomplianceandsecurityrulesoftheorganization.Anotherwayofdealingwiththeriskistograntaccesstoonlythosedatasourcesthatsomeonereallyneedstodohisorherjob.

Systems

Bysystemswemeanthephysicalandtechnicalenvironmentwheredataarestored.Thesedaysthe“cloud”isboomingandmoreandmorefirmsareputtingcrucialdatasourcesinthe cloud—which means that it is sometimes not clear where the data are physicallylocated. Especially due to legal implications (i.e. under which law the data should betreated)firmsshouldbeawareofwheretheirdataareatthegeographicallevel.Inmany

firms there are many systems for which data are stored, sometimes even redundantly,meaning that all these systems should be in scope when considering the necessarymeasuresfordatasecurity.Everysystemwillhaveitsowncriteriaonhowcriticalthedatain that system are for the daily operations of the firm, and what that implies for datasecurity. Typical measures regarding systems are measures such as physical access tosystems and computers (e.g. protocols for entering the rooms where the systems arelocated)orhowoftenandwhenbackupsaremade,andforhowlongitisacceptablethatasystemis“down.”

Processes

Wedefinethreetypesofprocesses:(1)processesforaccess;(2)processesforcontinuity;and(3)processes for steeringandmonitoring.Processes foraccessdefinewhenandwhohasaccess to what data and for which purposes. This means defining, for each user, thedifferent rights to work with or use data, and ensuring that every user has a strongusername/password combination that will be updated at prescribed times. Processes forcontinuity define the firm’s policies on how to dealwith calamities, how andwhere tobackup,whatfallbackoptionsareavailable,andhavingadisasterrecoveryplaninplace.The last processes are about steering andmonitoring. These processesmake sure that adatasecuritybaselineisinplacewithintheorganization,definingthecriticalperformanceindicators, including the standards for these indicators. Reporting on these indicatorsmakessurethatthesecuritymeasuresaremonitoredandthatthefirmisawareofpossibleincidents,includingtheactionstaken.

Conclusions

Privacyandsecurityhavebecomeimportantissueswithinbigdataanalytics.Privacy,inasense,directlylinksanalyticstothecustomer.Henceitisveryimportantwhenexecutingbigdataanalyticstotakeprivacyissuesintoaccount.Inthischapter,wehavetriedtogivea comprehensive overview of what privacy actually is and why it is important. Wediscussed some important privacy policies. We also considered how privacy actuallyimpactsanalyticsandmentionedspecificsolutionstocircumventthis.Finally,wediscussedsome important issues surrounding data security. In sum, we believe that privacy andsecurityshouldbeprimaryissueswhendoingbigdataanalytics. It isno longeronlythelaw department of the firm that shouldworry about this. Privacy and security actuallyimpactmarketing,marketinganalytics,andeventheboard.

Notes

1ThischapterisbasedonthereportoftheCustomerInsightsCenterbyBekeandVerhoef(2015).

2 See www.bcgperspectives.com/content/articles/digital_economy_consumer_insight_value_of_our_digital_identity/

(accessedSeptember18,2015).

3Seewww.dlapiperdataprotection.com(accessedSeptember17,2015).

4 See www.ftc.gov/news-events/press-releases/2010/12/ftc-staff-issues-privacy-report-offers-framework-consumers

(accessedSeptember17,2015).Thereportcanbedownloadedfromthiswebsite.

5WerefertoDeBruin(2015)foramoreextensivediscussiononethicaldecisionmaking.

References

Ackerman,M.S.,Cranor,L.F.,&Reagle,J.(1999).Privacyine-commerce:Examininguserscenariosandprivacypreferences.ACMConferenceonElectronicCommerce,1–8.

Ascarza, E.,&Hardie, B.G. S. (2013).A jointmodel of usage and churn in contractualsettings.MarketingScience,32(4),570–590.

Beke, F. T., & Verhoef, P. C. (2015). Privacy: Bedreigingen en kansen voor bedrijven enconsumenten. Report of Customer Insights Center (RUGCIC-2015-01), University ofGroningen.

DeBruin,B.(2015).Ethicsandtheglobalfinancialcrisis:Whyincompetenceisworsethangreed.Cambridge,UK:CambridgeUniversityPress.

Fletcher, K. (2003). Consumer power and privacy: the changing nature of CRM.InternationalJournalofAdvertising,22(2),249–272.

FTC(2012).Protectingconsumerprivacyinaneraofrapidchange:Recommendationsforbusinessesandpolicymakers.RetrievedfromFTC.gov.RetrievedSeptember18,2015fromwww.ftc.gov/sites/default/files/documents/reports/federal-trade-commission-report-protecting-consumer-privacy-era-rapid-change-recommendations/120326privacyreport.pdf

Godin, S. (1999). Permission marketing: Turning strangers into friends and friends intocustomers.NewYork:SimonandSchuster.

Goodwin,C. (1991).Privacy:Recognitionofaconsumerright.JournalofPublicPolicy&Marketing,10(1),149–166.

Holtrop,N.,Wieringa,J.E.,Gijsenberg,M.J.,&Verhoef,P.C.(2014).Nofuturewithoutthepast?Predictingcustomerchurnwith limitedpastdata.WorkingPaper,UniversityofGronigen.

Hong, W., & Thong, J. Y. L. (2013). Internet privacy concerns: An integratedconceptualizationandfourempiricalstudies.MISQuarterly,37(1),275–298.

Jones, K. (2003). Privacy:What’s different now? Interdisciplinary Science Reviews, 28(4),287–292.

MarketResponse(2014).Privacyhooginhetvaandel,MarketResponse,Amersfoort.RetrievedSeptember8,2015fromhttps://ddma.nl/kennisbank/privacy-hoog-het-vaandel/

Prins,R.,&Verhoef,P.C.(2007).Marketingcommunicationdriversofadoptiontimingofanewe-serviceamongexistingcustomers.JournalofMarketing,71(2),169–183.

TheBostonConsultingGroup.(2012).Thevalueofourdigitalidentity.Retrievedfrom

slideshare.net.RetrievedSeptember11,2015fromwww.slideshare.net/fred.zimny/boston-consulting-group-the-valueofourdigitalidentity.

Tucker,C.(2014).Socialnetworks,personalizedadvertising,andprivacycontrols.JournalofMarketingResearch,51(5),546–562.

Van Doorn, J., Verhoef, P. C., & Bijmolt, T. H. A. (2007). The importance of non-linearrelationshipsbetweenattitudeandbehaviourinpolicyresearch.JournalofConsumerPolicy,30(2),75–90.

Warren, S.D.,&Brandeis, L.D. (1890). The right to privacy.HarvardLawReview, 4(5),193–220.

Westin,A.F.(1967).Privacyandfreedom.NewYork:Atheneum.

Youn, S. (2009). Determinants of online privacy concern and its influence on privacyprotection behaviors among young adolescents. Journal of Consumer Affairs, 43(3),389–418.

4Howbigdataarechanginganalytics

Introduction

Analytics is a major element of creating value from big data. Statistical analytics ofmarketing data have been around for decades. The revolutions of scanner data andcustomerrelationshipmanagement(CRM)haveconsiderablyincreasedtheimportanceofanalyticsinmarketing:itcreatesstrongmarketandcustomerinsightsandmodelsthatcanbe used for decision support, campaigns, and information-based products. However, theemergingpresenceofbigdataischanginganalytics.Takingamorehistoricallens,wecanobservecertaindevelopmentsinanalytics.Wewillfirstdescribetheroleofanalyticsandgeneral types ofmarketing analysis. Subsequently,wediscuss thedifferent strategies foranalyzing big data.We endwith a discussion on how big data is changing analytics inspecificmarketingdecisionareasandingeneral(generictrends).Inthischapterwedonotdiscuss details of specific analytical techniques—wedo that in the in-depthChapters 4.1and4.2.Wediscusshow tohaveagreater impactwithanalytical results, through story-tellingandvisualization,inthein-depthChapter4.3.

Thepowerofanalytics

Inaneraofbigdata,firmsheavilyrelyontheanalyticalfunction.DavenportandHarris(2007) argue that firms can gain a competitive advantage if they build up strong andeffective analytical capabilities: “Analytics is then defined as the extensive use of data,statistical and quantitative analysis, explanatory and predictive models, and fact-basedmanagementtodrivedecisionsandactions”(Davenport&Harris,2007).Theseanalyticalcapabilities can be used in different kinds of functions, such as human resourcemanagement, logistics, finance, and marketing. The use of these capabilities may, forexample,leadtolesswasteinallkindsofprocessesandmayoptimizecertaindecisions.Forexample, retailersmay create an assortment in such away that category profitability isoptimized(e.g.VanNierop,Fok,&Franses,2008).Severalfirms,suchasAnnheuser-Busch,Google,Tesco,Wall-Mart,Fed-ExandHarrah’sEntertainment,haveadoptedstrategiesthatrelyheavilyon theanalytical function.DavenportandHarris (2007)argue thatcustomeranalytics is one of the more important fields where firms can compete. For example,CapitalOne has achieved growth through amore effective analytical-based targeting ofnewandcurrentcustomers.Theperformanceimplicationsofstrongermarketinganalyticshave been shown in several studies (e.g. Hoekstra & Verhoef, 2011; Germann, Lilien, &Rangaswamy, 2012).The positive and significant effects of analytical capabilities onperformance have been reported inmany industries (Germann, Lilien, Fiedler, &Kraus,2014),asshowninFigure4.1.Thelargesteffectsarefoundwithinretail:thesmallestwithinthebankingandsecuritiessector.

Figure4.1Associationsbetweencustomeranalyticsdeploymentandperformanceperindustry

Source:AdaptedfromGermannetal.(2014:591)

Differentsophisticationlevels

Davenport andHarris (2007) also distinguish several sophistication levels of analytics.Ahighersophisticationlevelshouldleadtolargercompetitiveadvantage(seeFigure4.2).

Onabroad level theydistinguishbetweenaccess, reporting,andanalytics.Accessandreporting is frequently a standardized tool—for example by online analytical processing(OLAP)—withinfirmsandismainlydescriptive innature.Onecouldalsorefer tothisasmarketorcustomerknowledge-focusedanalysis.Thisanalysis focusesongainingmarketandcustomerinsightsfromdescriptivestatisticalanalyses.Themainobjectivesaretolearnmore about markets and customers and to provide management with information onmarketingandcustomermetrics,suchasmarketshares,brandawareness,retentionrates,and customer profitability. These types of analyses focus on what has happened (past)instead of what will or could happen (future). Usually, simple statistical techniques areused, such as calculating averages. Access and reporting is pretty standard and is oftenincluded in management dashboards provided by software suppliers such as BusinessObjects, Cognos, etc. Analytics involves more sophisticated statistical techniques andanswersmorecomplicatedbusinessproblems. Italso focusesmoreon the future thanonthepastandismoreprescriptiveinnature.Analysesheremayanswerquestionssuchas,“What is the optimal product assortment to offer in a store?” or “What would be theoptimalprice level?” (seeFigure4.3). Yet other questions could be, “What is the optimalnumber ofmailers to be sent to a customer?” (e.g. Rust&Verhoef, 2005) and “Throughwhich channels should we contact a customer in order to optimize customer value?”According toDavenportandHarris (2007) firmsusing these typesofanalytics shouldbethewinnersintomorrow’sbusinesslandscape.

Figure4.2Differentlevelsofstatisticalsophistication

Source:AdaptedfromDavenportandHarris(2007)

Figure4.3Optimizationofmarketsharevs.revenueperpricelevel

Generaltypesofmarketinganalysis

Figure4.4Classificationofanalysistypes

Weclassifymoresophisticatedanalysesgoingbeyondthemarketandcustomerknowledgeanalysesontwodimensions(seeFigure4.4):

1. Explanatoryvs.predictiveanalyses2. Staticvs.dynamicanalyses.

The distinction between explanatory and predictive analyses is quitewell known and isratherimportant.Explanatoryanalysesfocusonthe“Why?”question.Aresearcheraimstoknowwhyspecificphenomenahappen.Forexample,whenrunningcustomeranalyses,onewouldliketoknowwhy,forexample,specificcustomersspendmore.Thenonecouldlookfordifferences between customerswith ahigh and lowprofitability.This couldbedonewithdifferentanalyseswithdiffering levelsofstatisticalcomplexityandrigor.Predictiveanalyses focus on forecasting marketing- and customer metrics. Market forecastingtypicallyfocusesonsalesforecastingofnewproducts,andforecastingmarketshare.Atthecustomer level thismayinvolveforecastingresponsetomarketingactions,suchasdirectmails,emails,butalsoforecastingchurn,productreturns,andlifetimevalue(e.g.Donkers,Verhoef,&DeJong,2007;Risselada,Verhoef,&Bijmolt,2010).Again,predictiveanalysescanbedonerathersimply,perhapsbyusingakindofweightedaveragetopredictfuturesales, or in a more sophisticated fashion by using forecast models, such as time-seriesmodelsorchoicemodels.Interestingly,simplemodelssometimesperformaswellasmorecomplicatedmodels.Donkers et al. (2007) show that a simplemodel topredict customerlifetimevalue(CLV)withpastprofitabilityasanestimateforfutureprofitabilitypredictsjustaswellasamuchmorecomplicatedchoicemodel.

Bothexplanatoryanalysesandpredictiveanalysescanbeexecutedinastaticwayoramoredynamicway.Withstaticanalyses,onetypicallyanalyzescross-sectionaldata.Thatis,ataspecificpointintime(t)onehasdataon,forexample,brandsorcustomers,andone

then analyzes these data. Specifically, one can then observe how brands differ or howcustomersdifferandwhytheydiffer.Onecanevendevelopapredictivemodeltopredictaspecificevent,suchaschurnatthatpointintimeusingtheavailabledata.Thismodelcanthenbeusedtopredictthiseventinthefuture.Staticmodelsdominatepractice,ascross-sectionaldataaremorecommon.Inadynamicanalysis,dataovertime(t,t-1,t-2,…t-n)areanalyzed.Asananalystonecanthenobservehowspecificmetricschangeovertimeandhowthiscanbeexplained,andwhetherthesechangescanbeforecast.Dynamicmodelsaregenerally preferred to gain insights into the effectiveness ofmarketingmetrics and howchangesintheenvironmentaffectmarketingoutcomes(Leeflangetal.,2009).Somerecentresearchhasputforwardanadvanceddynamicmodelthatholdsoutthepromiseofstrongpredictive power, and takes into account differences between customers (e.g. Holtrop,Wieringa,Gijsenberg,&Verhoef,2014).

Strategiesforanalyzingbigdata

The presence of big data provides huge opportunities for analytical teams. One of theeasiestwaysoftakingadvantageofbigdataisprobablyjusttostartupanalysesandstartdigging into theavailabledata.Bydigging into thedataonemightgainvery interestinginsights,whichcanguidemarketingdecisions.ThemostfamousexampleinthisrespectisthatofUK-basedretailerTesco:whenanalyzingdataoftheirloyaltycard,theydiscoveredthat consumers buying diapers also frequently buy beer and chips (Humby, Hunt, &Philips,2008).Althoughsuchanexamplecanbeinspiring,wepositthatbeforestartingupananalyticalexerciseoneshouldclearlyunderstandthebenefitsanddisadvantagesofthespecificanalysisstrategy—aswellasthatofotherstrategies.

In our two-by-two matrix we distinguish between four basic analysis strategies (seeFigure4.5).Wetake intoaccounttwodimensions.First,analysescanbestartedbasedonwhetherornotaproblemispre-defined.Apre-definedproblemcanarisefrom:

Marketing challenges (e.g. decreasing loyalty, eroding prices, lower acquisitionrates)Marketinggrowthobjectives (e.g. achieve salesgrowthof 20%, improve customersatisfaction).

Analytics can be done on the pre-defined data (e.g. a CRM database on customertransactions),butonemayalsoaimtolookforavailabledataandcombinethesedatabasedon the arising needs in the data analysis. This is the second dimension of our analysisstrategy framework. Based on these two dimensions we distinguish four analyticalstrategies:

1. Problemsolving2. Datamodeling3. Datamining4. Collateralcatch

Figure4.5Bigdataanalysisstrategies

Problemsolving

From a scientific perspective a problem-solving analytical strategy is deductive.Usually,theanalyststartswithamanagerialproblemorissue.Problemsmayinvolve:“Howcanweincrease the value of our customers?” or “How canwe improve the net promoter score(NPS)?” or “Which pricing strategy should I use to attract more profitable customers?”After defining the problem hypotheses, assumptions could be defined, explicitly orimplicitly,aboutpotentialsolutionsoftheproblem.Forexample,whenstudyingdriversofNPS, analysts coulddevelopa list ofpotentialdeterminantsofNPS, suchasadvertising,socialmediamessages,theserviceexperience,etc.Thespecificprocess,fromdefiningtheproblemtodevelopingalistofhypotheses,guidestheselectionofdatatobeanalyzed(seeFigure4.6).FollowinguponthedriversofNPS,forexample,researcherswouldprobablylook for data that includes NPS and combine that with data on potential drivers, oralternatively researchersmight collectnewdata.Abigdifferencebetween this approachandthedataminingapproachisthatresearchersdonotstartwiththedata.Bigdatacanbeusedtosolvetheproblem,butitisnotanecessitytosolvetheproblem.Theproblemcanalsobesolvedwithlimiteddatadrawnfromexistingavailabledata(e.g.aCRMdatabase)orresultingfromanewdatacollectioneffort(e.g.asurvey).

Datamodeling

Figure4.6Problem-solvingprocess

There is also a problem that needs to be solved in the data modeling approach. Forexample,oneobjectivecouldbetopredictchurn(seeFigure4.7).Thedifferencewiththeproblem-solvingapproach is that the focus ismoreondata and especiallyon theuseofnew data sources. Using new data sources one could, for example, aim to find newpotentialpredictorsofchurn.Onepotentialpitfallwiththisapproachisthatanalystsfocustoomuchonthedataandlackstrongconceptualizationsonwhyspecificrelationshipsinthedataarefound.Outcomesmayeasilybespuriouscorrelationsinwhichtheassociationisbasedonanunderlying,unobservedvariable.Theapproachmayalsoleadtoundesirableoutcomes, suchas the famousexample inwhicha fatherofa teenagerwasofferedbabyproducts, because his apparently pregnant daughter purchased baby products using hisloyaltyprogram.Thisisclearlyrelevanttotheprivacyandbehavioraltargetingdiscussionput forward in Chapter 3.3. The approach can become overly data-driven instead ofproblem-driven, easily turning into a data-mining exercise.However, an advantage over

theproblem-solvingapproachisthatthedata-modelingapproachismoreflexibleintermsofdatausage.Thismayresultinmoreinnovativemodelsolutions.

Figure4.7Churnmodelresultsfortelecomfirm

Source:AdaptedfromRisseladaetal.(2010)

Datamining

Thisisamuchmoreexplorativeanalyticalstrategy.Typicallydataarenotpre-definedanda defined problem does not guide the analysis. Also, no hypotheses are implicitly orexplicitly stated. From a scientific perspective it is a typical inductive analysis. The keybelief is that thewidelyavailabledatacanprovidevaluable insights justbydigging intothedata.Indoingso,relevantnewrelationshipscanbediscoveredthatcanpotentiallybevaluable.Theclassicdataminingexample,asmentionedearlier,isthatwhenanalyzingtheTescoClubCarddata, analystsdiscovered that customersbuyingdiapersalso frequentlypurchasedbeerandchipsonaFridaynight(seeFigure4.8).Thisfindingcouldbeusedtotargetpromotions.Thediscoveryof thesepatterns can create innovations.However, onemajor potential pitfall is that the analyses are unguided andmay result in all kinds ofassociationsthataredifficulttointerpret.Furthermore,asnospecificproblemsaresolved,

manyoftheseanalysesareoflittleuse,andmaynotofferanyimpact.

Collateralcatch

Thelastanalyticalstrategyisnotanexplicitstrategy.Itisratheraby-productofaproblemfocusedanalysis.Whenanalyzingpre-defineddata,analystsmaysometimesdiscovernewrelationships, which can be very valuable. For example, when analyzing the impact ofdifferent touchpoints on conversion rates for an online retailer,we found differences inconversion rates for different devices (e.g. mobile, tablet, desktop), which we were notlooking for inouranalysis (seeFigure4.9).This inducedus to study conversion rates inmoredepth, and specificallyhowdevice switching in thepurchase funnelwould impactconversion rates. These relationships are called “collateral catches” because they are notsought initially, and only become available as a result of some unusual feature of theresearch,orthroughsomeexploratoryanalysisbeingexecutednexttothemorestructuredproblem-solving analysis.One thingwe have learned is that analysts should be open tothese collateral catches. They may provide a deeper understanding of the phenomenonstudied.Furthermore,thesecatchesmaybevaluableinguidingnewdirectionsforsolvingthedefinedproblem,orinprovidingnewinnovationsforexecutingmarketing.

Figure4.8Tesco’sbeeranddiaperdata

Figure4.9Differentconversionratesafterdeviceswitching

Source:AdaptedfromDeHaan,Kannan,Verhoef,&Wiesel(2015)

Howbigdatachangesanalytics

Thegrowthofbigdataischanginghowanalyseswillbeexecuted.Itwillalsochangethescopeofanalyticalquestionstobeanswered,asmoredataareavailable.Inourdiscussionweagain focuson threeareasofmarketinganalytics: (1)market; (2)brandandproduct;and(3)customers.Wehavenoticedthattheseareasare,asalreadyshown,interrelated,butourexperienceisthattheyaretreateddifferently.Moreover,someindustries—forexample,fast-movingconsumergoods(FMCGs)—usuallyhaveatypicalbrand/productfocusduetothe nature of their markets and product/brand offerings, while other industries (e.g.financial services, telecoms) have a stronger customer focus in their analytics. In theselatter industries, firmsalso typicallyhavemorecustomerdata (CRMdatabases) than theproduct/brand-focused industries. Market analytics are equally relevant for all firms, astheyfocusonhowmarketsdevelop,andwhatthatimpliesforfirmstrategies—forexample,for allocation of resources across strategic business units (SBUs), brand and productmanagementdecisions,andcustomerstrategies.

AsalreadydiscussedinChapter3oneofthemajorchallengeswithbigdataisthatdataischanging(seeforexamplethe3Vmodel).Averyimportantdevelopmentwithregardtodata is thatwearemoving fromsingle sourcedata tomultiple sourcesofdata.Thishasconsiderableconsequences,aswithmulti-sourcedatayouhavetoconnectdatainasmartway. This is not always obvious: connecting can seem easy but one needs a commonvariable that can be linked (see Chapter 3.1). Connecting data may also involve dataconnectingatmultiplelevels.Forexample,customerdatacanbelinkedtotime-seriesdatasuchasthatarisingfromadvertising(Prins&Verhoef,2007);onethenhasdataatdifferentaggregation levels. The challenge here is that there is much variation in data betweencustomers,butthatthevariationoftime(i.e.inadvertisingordistribution)maybelimited,whilethenumberofdatapointsatthatlevelisalsosmallcomparedtothenumberofdatapoints at the customer level. Thismay also induce theuse ofmore complicatedmodels,suchasmulti-levelmodels.Anotherimportantdevelopmentisthatdataisbecomingmoreunstructured,with text data being themost important example. But also social networkdata can be included, especially at the customer level. These data are also inherentlycomplex(e.g.Risselada,Verhoef,&Bijmolt,2014).

Weshowhowbigdata is changing thedata,analytics,and the relevance/scopeof theanalytics in Figure 4.10. Based on this schematic overview we discuss some of thesedevelopments,focusingmainlyonchangesinanalytics.

Marketlevelchanges

Figure4.10Howbigdataarechanginganalytics

Atthemarketlevel,timeseriesanalysesarethemostcommonwayofforecastingmarketdevelopments.However,inthepresenceofbigdata,weobservethatfirmswanttoassesstheimpactofspecificchangesintheenvironment.Thisisnotimpossiblewithtraditionaltime-seriesanalysis,asinthesemodelsexplanatoryXvariablescanbeincluded,inmodelssuchasVARXandARMAX(Franses&VanDijk,2000).However,thedominatingimpactoflaggedvariablesand/ortrendeffectsmightreducetheeffectsofexplanatoryvariablessuchas a changing environment. Moreover, at the aggregated level, environmental changesmightnotbeobserved,giventhat itonly impactsaspecificcustomersegment.Thismayespecially hold for disruptive changes (Sood&Tellis, 2011).More specifically, disruptiveinnovations may be targeted at a relatively small segment in the market, but may beadoptedbythemajorityofthemarketwhenittakesoff.Theearlyadoptionandincreasingrelevanceofthisinnovationislikelynotobservedwhenanalyzingaggregate-leveldatajustbecausethechangeistoosmallandthetrendisstillupwards.Toillustratethiswediscussthe caseof theDutch telecom industry around2010.The telecommarkethaddevelopedstrongly in the previous decade, but as the adoption ofmobile telephoneswas reachingmaturity, growth was declining. Moreover, the adoption of smartphones such as theiPhone,Blackberry, and theSamsungGalaxy increased theusageof Internet services onmobiles.Oneoftherevenuedriversfortelecomfirmswastheusageoftextmessages(i.e.SMS). Customers frequently used more text messages than their contract provided for,resulting inhigherrevenuesasusageabovethecontract levelswaspricedhigher.Hence,thisservicewasreallyacashcow.In2010anewindependentserviceenteredthemarket:WhatsApp.Throughanapponthesmartphone,customerscouldchatwitheachotherforfree.Thefeeforusingthisappwaszerointhefirstyearandafterthatitcostaround0.70euro per year. The app was first adopted mainly by young customers. These youngcustomersmainlyusedabrandfocusingonyoungconsumers.Ontheaggregate levelnochangesintextusagewerefound.However,strongchangescouldbefoundinasmall,butinfluentialsegmentinthemarket.ThediffusionofWhatsAppwasveryrapid,andafterafewmonthsaggregatedsalesfiguresshoweddecreasesintextusage(seeFigure4.11).Thesechanges could, however, not have been predicted with traditional time-series analysis.

Moredeep-divingintosegment-leveldata

Figure4.11ImpactofWhatsAppusageonthesmartphoneusageofaDutchtelecomcompany

was required to forecast these changes. Unfortunately in many telecom providers, themyopictopmanagementfoundithardtobelievethatthesechangesatthesegment levelcould actually finally affect the whole customer base. Studying text analyses of onlineforumdiscussionscouldprobablyhavehelpedtoreallyunderstandwhatwashappening.1

Brandandproductchanges

Brand and product management analytics traditionally have been rather descriptive,mainlyusing surveydataonbrandawareness and/orpanel level salesdata.These latterdataweremainlyavailableataweeklylevelforFMCGsasaconsequenceofthescannerdata revolution. Primarily in the academic community, researchers have considered theimpact of marketing mix instruments on sales and/or market share (Leeflang, Wittink,Wedel, & Naert, 2000; Nijs, Dekimpe, Steenkamp, & Hanssens, 2001). Similarly, marketresearch agencies have developedmodels such as ScanPro by ACNielsen, to assess theimpactofpromotionsonbrandsales.Thebigdatadevelopmentis,however,alsoaffectingbrand and productmanagement. Specifically, the use ofmulti-source data allows brandmanagers to assess the impact of themarketingmixonattitudinal brandmetrics (brandshare of mind metrics, i.e. preferences) and their simultaneous impact on sales (e.g.Hanssens,Pauwels,Srinivasan,Vanhuele,&Yildirim,2014).

The changing nature of big data analytics for brands becomes rather clear in thefollowingexample.Asachallenger,aDutchtelecomoperatorhasgainedaseriousmarketshareinahighlycompetitivemarketoveraperiodoftwoyears.Inordertorealizefurthergrowth,theywantedtore-positiontheirbrand.Thecompanythereforehasastrongneedfor a fact-based approach to underpin theirmarket development strategy.An importantinsight that made them feel comfortable investing in their brand communication wasgained from a multi-source data analysis that showed a positive relation between

development of their brand performance (in terms of awareness, consideration, andpreference) and their currentmarket position in terms of sales share/market penetration(seeFigure4.12).

Anotherdevelopmentisthatdigitaltextanalyticsresultsinnewmetrics,suchasdigitalbrand sentiment indices,which canbemeasuredover timewithouthaving any surveys.Thesemetricscanbelinkedtosales,andalsotostockprices(e.g.Tirunillai&Tellis,2012).In specific markets, such as movies, these metrics seem highly relevant for predictingproductsuccess(e.g.Onishi&Manchanda,2012).Ingeneralweexpectbigdatatoallowfora better assessment of several accountability issues: the effect of investment in brands,shareofmindbrandmetrics,andmarketperformance.Newdatacollectionmethods,suchas amobile survey,may be used to understand the impact of different touch-points onbrand preference (Macdonald,Wilson,&Konuş, 2012;Baxendale,Macdonald,&Wilson,2015).

Figure4.12Caseexampleofmulti-sourcedataanalysisofrelationbetweenbrandperformanceandsalesshare

Customerlevelchanges

Major changes are expected at the customer level. Compared with market level andbrand/product level data, customer data have already been relatively rich due to thedevelopments in CRM. This increasing presence of data has especially accelerateddevelopmentsinindividual-levelmodelsdesignedtopredictindividual-levelbehaviors(e.g.Blattberg,Kim,&Neslin,2008;Risseladaetal.,2010).Italsostirredupthedevelopmentofnewmetrics,suchasCLV(Venkatesan&Kumar,2004).Thebigdatarevolutionhaspushedsinglesourcedataofindividualcustomerstooneside,tobereplacedbymulti-sourcedata.Hence, individual data will be enriched with other sources of data. For example, for aEuropean public transport company we enriched individual-level survey data onsatisfactionwithoperationsdata to fullyunderstand thedriversofcustomersatisfaction.This provided interesting insights. For example, operations data on punctuality are animportantdriverofsatisfaction(Gijsenberg,VanHeerde,&Verhoef,2015).

Wehavealreadymentionedthatindividual-leveldatacanbecombinedwithaggregated

level marketing data. Similarly social network data can be added. One of the majordifficultiesofthisisthatcustomersinteractatdifferenttouchpointsonlineandoffline,andsimilarly with online and offline advertising media. Big data analytics are required foronline firms aiming to assess the importance of different online and offline advertisinginvestmentsononlineconversionrates(e.g.DeHaan,Wiesel,&Pauwels,2013).

Another major change is the extension of the customer management paradigm, byfocusingmore on customer engagement. Customer engagement goes beyond traditionalcustomertransactions;italsofocusesonnon-transactionalbehaviorofcustomers,suchasword-of-mouth, likesonsocialmedia,blogpostsetc. (e.g.VanDoornetal.,2010).Hencemodelsnowadaysincludesocialmediadata,suchasindividualcustomeronlinepostsandlikes,toaccountforcustomerengagementactivities(DeVries,Gensler,&Leeflang,2012).Theincreasingpresenceofcommunitiesalso inducesastrongerfocusonsocialnetworksandtheimportanceofthesenetworksincreatingvalueforfirms(e.g.Hennig-Thurauetal.,2010;Libaietal.,2010).Thiscreatesastrongerdemandforsocialnetworkanalyses.

Genericbigdatachangesinanalytics

Asdiscussed inChapter2, big data is frequently characterizedby the threeVs:Volume,Velocity, and Variety. Sometimes two other Vs are added: Veracity and Value. Thesechanges indatamaysuggest thatanalyticswillchangedramatically,and it is sometimeseven referred toaskindofdisruptiveevent (e.g.Sathi, 2012).Wehavealreadydiscussedsomechanges, suchas themovement tomoreunstructureddata, in theprevioussection.But there are some additional changes, and our views on these changes are describedbelow.

Fromanalyzingsamplestoanalyzingthefullpopulation

The increasingvolumeofdata in a bigdata era suggests thatwenowcananalyzeverylargedatabaseswithmillionsofobservations.Wearemovingfromstudyingasmallsampleto studying the full population. The rising computer power of the last few years hasfacilitated this,while alsodata storage capacity seemsunlimitedwith sufficient space inthe cloud. Indeed, it sounds very cool and convincingwhen an analyst can report thatmillions of observations have been analyzed. The advantages of analyzing very largedatabasesis,however,relativelysmall.Actually,theoutcomesofanalysisofatruerandomsampleofobservationsofapopulationofmillionsshouldnotdiffersubstantiallyfromtheanalysis of the full population. It is not the volume that is important, it is importantwhether theanalyzed sample is representativeof thepopulation studied.Volume isonlyimportantinordertogetmorereliableanswersatahighersignificancelevel.Ananalystshould, however, be more concerned about biases due to a wrong representation ofcustomers,brandsetc.intheirdata,ratherthanthevolumeofthedata.Thisisnottosaythatthesizeoftheanalyzedsampleisnotrelevant.Increasesinlowernumbersofsamplesize (e.g.moving from400 to2,000observations)canbeespeciallyvaluable,as reliabilitymayincreaseandalsospecificeconometricissues,suchascollinearindependentvariables,maybecome less of an issue.However,moving from20,000 to 50,000 observations or to100,000 observations will become less rewarding for analytical purposes. In general weprovidethefollowingsimplerulesforwhenlargersamplesbecomemorevaluable:

If more variables are studied in an analysis and specifically when studying theeffectsofmultiplevariablesonanoutcomevariableIfoneneedstostudydifferentsub-samples.Thesub-sampleshouldbeofsufficientsizetoanalyzeIfthereismuchheterogeneity.Forexample,customersmaydifferstronglyinhowtheybehave,andhowtheyrespondtomarketingactionsIf the studiedvariableoccurs very rarely (e.g. conversiononan email campaign)andasufficientnumberofdatapointsisrequiredtounderstandthedriversofthis

eventIfthereisstrongcollinearitybetweentheindependentvariablesusedtopredictorexplainadependentvariable,suchassalesorchurn.

Importantly,althoughthevolumeofavailabledataisindeedincreasing,weobservethatitisnothappeningforalltypesofdata.Attheindividualcustomerlevel,datavolumeshaveindeed become huge and can be enriched with many others. Especially in the onlineenvironment,datacanbecomemassive.However,brand-leveldataon,forexample,brandsales, brand preferences, and advertising are frequently still limited. For example, for aEuropean public transport company we analyzed the effects of advertising on travelledkilometerspermonth,andwhilehaving threeyearsofdata thisonlyresulted in31datapoints(Gijsenberg&Verhoef,2015).

Fromsignificancetosubstantiveandsizeeffects

Analyzinglargesamplesfrequentlyleadstomanyhighlysignificanteffects.Smallp-valuesseem the norm rather than the exception. However, in a big data era with these largesamples,wearguethatweshouldmoveawayfromsignificanceandfocusmoreonthesizeof theseeffects.Witheffect sizewe look forwhether the foundeffectsaresubstantialorsmall. For example, theremight be a difference of one year in the average age betweenswitching and loyal customers (e.g. 43 vs. 42 years),which is highly significant in a bigdataanalysis.However,onemightquestionthesizeofthisdifference.Shouldafirmthenfocusmoreonyoungercustomerstopreventswitching,asthesecustomersarelessloyal?Similarly, the addition of variables in an explanatory model will certainly frequentlyincreasetheexplainedvarianceofamodel.Thekeyfocusshouldthenbeonthesizeofthisadditionalexplanatorypower—isitreallysubstantialorisitonlyincremental?Further,itisnotonlythesizeoftheeffectthatshouldmatter,butalsothesubstantiveandmanagerialmeaning of a found effect. Going back to the loyalty example, one might question themanagerialimplicationofsuchaneffect.Shouldafirmdevelopdifferentspecificstrategiesfor younger customers than they do for older customers, as young customers aremorelikely to switch? We have some specific recommendations for actions that an analystshouldtakewheninterpretingbigdataresults:

FocusonthesizeofthesignificanteffectsinsteadofsignificanceonlyVisualizethefoundeffectsandconsidertheeffectsizesCompare the found effects with existing benchmarks (e.g. when assessing theeffectiveness of socialmedia, one could compare itwith the effect of traditionaladvertising)Developmarketingimplicationsforfoundeffectsandchallengethem.

Fromadhocdatacollectiontocontinuousdatacollection

Figure4.13Differenttypesofdataapproaches

Wehavemovedfromadhocprojectswithdatatocontinuousdatacollection.Weconsiderfourtypesofresearch-baseddatadesignsontwocharacteristics(seeFigure4.13):

1. Repetitionsofrespondents2. Repetitionofmeasures.

Adhocdata projects frequently involve surveydata collection efforts to answer specificquestionssuchas, “Whatarecustomers looking for inanewproduct?”or“Howcanmybrandbestbepositioned?”Adhocdataforspecificissuescanalsobecollectedamongthesamesampleofrespondentsusingapanel.Thiswillfrequentlyoccurifafirmworkswithamarketingresearchagencyapplyinganonlinepanel,andifitdoeshappen,theremightbeopportunitiesforabigdataanalysttolinkthespecificdatacollectionefforts.Whatwenowobserve very frequently is data collection inwhich different samples are used to collectthese data, which typically will be data on value-to-customer (V2C) metrics. This caninvolvemonthly trackingofbrandevaluationsusinganonlinemarketing researchpaneland continuous survey research on satisfaction and NPS and other customer feedbackmetrics. In these kinds of studies different customers provide answers on the samemeasure, which allows firms to track their performance on these metrics. However,customerscannotbefollowedovertime,andthisthereforeleadstorepeatedcross-sections.Theaverageperformanceofthesestudiescanbelinkedtoinputvariables,suchasserviceperformance affecting service quality perceptions (Gijsenberg et al., 2015) or brandpreference metrics explaining brand sales (Srinivasan, Vanhuele, & Pauwels, 2010).However,customersarenoteasilyfollowed.Heresmartdatafusiontechniques, inwhichlook-a-likesarebeinglookedforinthedifferentsamples,mightbeasolution.Duetothelargernumberofindividualsinasegment,thesevariationscancelout,sotospeak,andtheobservedpatternsbecomemore reliable.Longitudinalpaneldata are readily available ingood customer databases in which customer behavior is tracked over time. Thelongitudinalbehaviorofvisitorscanalsobefollowed,inanonlineenvironment,byusingcookies.LongitudinalmeasuringofV2Cmetricsamongthesamegroupofrespondentsispossible, but a huge challenge, as customerswill drop out just bynot responding to the

next survey (e.g.Verhoef, 2003).Longitudinaldatado,however,have theadvantage thatonecanstudyhowchangesatthecustomerlevelaffecttheirbehavior(Verhoef,Franses&Donkers,2002).Wehavethefollowingrecommendations:

Setupcontinuousresearchonkey(V2C)metricsBeawarethatcross-sectionsdohavesimilarsamplecharacteristicsovertimeBe consistent in themeasurement of your importantmetrics.Consistency allowsthebigdataanalyststoreportandanalyzethesemetricsovertimeAlongitudinalpanelisnottheholygrail!Repeatedcross-sectionscanworkrathereffectivelyformanypurposes.

Fromstandardtocomputersciencemodels

ThedevelopmentofbigdataispartiallyembracedbyITandcomputerscience.Thisisofcourse not illogical, as IT developments have been a driving force behind big datadevelopments.Computerscientistshavedevelopedvariousmethodsofanalyzingdataoverthe last few decades: one of themost famous early exampleswas the so-called “neuralnetworks.”Morerecently,differentmachine-learningtechniqueshavebeenadded,someofwhichhavegainedattentioninthemarketingliterature.Probablythemostimportantoneis the“bagging-boosting” technique,whichestimatesspecificmorestandardmodelsonalargenumberofsub-samples,aimingtocomeupwiththebestprediction(e.gLemmens&Croux,2006;Risseladaetal.,2010).Oneofthemainproblemsthatmanypeoplefacewithcomputer-science-basedmodelsisthatitisfrequentlyconsidereda“blackbox.”Thatis,itis not clear to analystswhat themodel actually does andhow it is specified.Models inmarketingtypicallyhaveaneconometricbackgroundandcanbespecified(e.g.Leeflangetal.,2000).Anadvantageofthesemodelsisthattheyaremorestronglybasedonex-antesetexpectationsofeffects,whereascomputersciencemodelsmayeasilyleadtodatamining,withoutastrongtheoreticalandpracticalbase.

Theevidenceofastrongerperformanceofcomputersciencemodelsismixed.Ingeneralitseemsdifficulttocomeupwiththebestmodelacrossapplications.Studyingaround14differentdatabases,Donkers,LemmensandVerhoef(2014)reportthatthereisnomethodthatisconsistentlywinning.Theprobabilitythatamethodperformsbestseemstohighlydependonthedatabeinganalyzed.Theythereforeproposethatthebestwaytogetgoodresultsistocombinethedifferentmethods,takingadvantageofthestrengthofeachofthemethods.Somerecommendationsforanalyststofollowwhenapplyingmodelsare:

Understandthebackgroundsandpitfallsof(new)modelsbeforeapplyingthemBe inherentlyskepticalabout thecommunicatedperformanceof (new)modelsbysoftwareprovidersTesttheperformanceofdifferentmethodsondatabeinganalyzed

Selectamethodthatcanbecommunicatedtomarketingmanagersandhasagoodperformance. That is, find optimal balance between performance andmanagerialinsightfulnessUse visual aids to communicate findings of specifically computer-science-basedmethods.

Fromadhocmodelstoreal-timemodels

Inanenvironmentwheredatabecomeeasilyavailable,wemovetoatrendwheremodelscan be updatedmuchmore frequently. The need formodel updates is clearly shown inextant research. Especially in more turbulent environments models can easily becomeoutdated and as a consequence their findings are no longer valid and their predictionqualitydecreases. For example, in a telecomsettingRisselada et al. (2010) show that thepredictiveperformanceofmodelsdecreasessubstantiallyinperiodsfurtherawayfromthetimewhenthemodelwasdeveloped(seeFigure4.14).

Luckily data is now available much more quickly, so that updating models is mucheasier. Especially in the online environment, data can be available in real time. Thisprovides opportunities for constantly updating models with new data in constantlychanging(online)environments.Thesemodelscanthenbeused inonline targeting.Thismove to constantly updated models probably won’t happen in offline environments.However,herealsoweadvocatemoremodelupdating,asitisunlikelythatmodelresultshold for a long time.Thedynamics of today’smarket,with frequently changingmarketenvironments and changing customer behaviors, require a constant updating ofmodels.Wehavesomerecommendations:

Assess the stability of a developed model and the predictive performance of amodelovertimeDonotexpectyourmodel-results,resultinginsightsandpredictiveperformancetohaveeternallifeIn online environments base models can be updated to real-time models fortargetingpurposeswithreal-timedataasinput.

Figure4.14Averagetop-decileliftsofmodelestimatedattime

Source:AdaptedfromRisseladaetal.(2010)

Conclusions

In this chapter we have focused on analytics. We specifically discussed the power ofanalytics. We strongly believe that analytics can improve the quality of marketingdecisionsandinducesmartermarketingdecisions.Analyticscanfocusonbuildinginsights,(predictive)models,andoptimizationmodels.Itisimportanttodistinguishbetweenstaticanddynamicanalyses,aswellasbetweenpredictiveandexplanatoryanalyses.Weclearlyhavelaidoutdifferentanalysismethods.Wehaveastrongpreferenceforapproachesthataremore problem-solving in nature. The big data environment is changing analytics ineverymarketing decision.We have discussed these changes and specifically focused onsomefrequentlystateddevelopments,suchastheneedforlargeranalysissamples.Inthein-depthchapterswewilldiscussexistinganalyticalclassics(Chapter4.1)andnewbigdataanalytics(Chapter4.2),whilewealsodescribehowstory-tellingandvisualizationcanbeeffectiveincreatinganalyticsthathavemoreimpact.

Note

1Moreaboutthisclassiccaseintelecomcanbereadinwww.telecompaper.com/pressrelease/dutch-consumers-turn-to-

mobile-apps-for-communication-needs-43-of-smartphone-users-has-whatsapp-installed—813359(retrievedonOctober

302015).

References

Baxendale,S.,Macdonald,E.K.,&Wilson,H.N.(2015).Impactofdifferenttouchpointsonbrandconsideration.JournalofRetailing,91(2),235–253.

Blattberg,R.C.,Kim,B.-D.,&Neslin,S.A. (2008).Databasemarketing–Analyzingandmanagingcustomers.NewYork:SpringerScience&BusinessMedia.

Davenport,T.,&Harris, J. (2007).Competingonanalytics–Thenewscienceofwinning.Boston,MA:HarvardBusinessSchoolPress.

De Haan, E., Kannan, P. K., Verhoef P. C., & Wiesel T. (2015), The impact of deviceswitchingonconversionrates.WorkingPaper,UniversityofGroningen.

DeHaan,E.,Wiesel,T.,&Pauwels,K.(2013).Whichadvertisingformsmakeadifferenceinonlinepathtopurchase?MSIworkingpaperseries,Boston.

DeVries,L.,Gensler,S.,&Leeflang,P.S.H.(2012).Popularityofbrandpostsonbrandfanpages:Aninvestigationoftheeffectsofsocialmediamarketing.JournalofInteractiveMarketing,26(2),83–91.

Donkers,B.,Lemmens,A.,&Verhoef,P.C.(2014).Thepredictivepowerofchurnmodels.WorkingPaper,ErasmusUniversityRotterdam.

Donkers,B.,Verhoef,P.C.,&DeJong,M.G.(2007).PredictingCLV:Atestofcompetingmodelsintheinsuranceindustry.QuantitativeMarketingandEconomics,5(2),163–190.

Franses,P.H.,&VanDijk,D. (2000).Non-linear timeseriesmodels inempirical finance.Cambridge:CambridgeUniversityPress.

Germann, F., Lilien, G. L., & Rangaswamy, A. (2012). Performance implications ofdeployingmarketinganalytics. International Journal ofResearch inMarketing,30(2),114–128.

Germann,F.,Lilien,G.,Fiedler,L.,&Kraus,M.(2014).Doretailersbenefitfromdeployingcustomeranalytics?JournalofRetailing,90(4),587–593.

Gijsenberg,M.J.,VanHeerde,H.J.,&Verhoef,P.C.(2015).Lossesloomlongerthangains:Modeling the impact of service crises on customer satisfaction over time. Journal ofMarketingResearch,52(5),642–656.

Gijsenberg, M. J., & Verhoef, P. C. (2015). Moving Forward: The Role of Marketing inFosteringPublicTransportUsage,WorkingPaperUniversityofGroningen.

Hanssens, D. M., Pauwels, K. H., Srinivasan, S., Vanhuele, M., & Yildirim, G. (2014).Consumer attitude metrics for guiding marketing mix decisions.Marketing Science,33(4),534–550.

Hennig-Thurau,T.,Malthouse,E.C.,Friege,C.,Gensler,S.,Lobschat,L.,&Rangaswamy,

A. (2010). The impact of new media on customer relationships. Journal of ServiceResearch,1(3),311–330.

Hoekstra,J.C.,&Verhoef,P.C.(2011).Thecustomerintelligence–marketinginterface:Itseffectonfirmperformance.WorkingPaper,UniversityofGroningen.

Holtrop,N.,Wieringa,J.E.,Gijsenberg,M.J.,&Verhoef,P.C.(2014).Nofuturewithoutthepast?Predictingcustomerchurnwith limitedpastdata.WorkingPaper,UniversityofGronigen.

Humby,C.,Hunt,T.,&Phillips,T.(2008).Scoringpoints:HowTescoiswinningcustomerloyalty.Philadelphia:KoganPagePublishers.

Leeflang,P.S.H.,Bijmolt,T.H.A.,Doorn,J.,Hanssens,D.M.,VanHeerde,H.J.,Verhoef,P.C.,&Wieringa,J.E.(2009).Creatingliftversusbuildingthebase:Currenttrendsinmarketingdynamics.InternationalJournalofResearchinMarketing,26(1),13–20.

Leeflang, P. S.H.,Wittink,D. R.,Wedel,M.,&Naert, P.A. (2000).BuildingModels forMarketingDecisions.Dordrecht,theNetherlands:KluwerAcademicPublishers.

Lemmens, A., & Croux, C. (2006). Bagging and boosting classification trees to predictchurn.JournalofMarketingResearch,43(2),276–286.

Libai,B.,Bolton,R.,Bügel,M.S.,DeRuyter,K.,Götz,O.,Risselada,H.,&Stephen,A.T.(2010). Customer-to-customer interactions: Broadening the scope of word of mouthresearch.JournalofServiceResearch,13(3),267–282.

Macdonald,E.K.,Wilson,H.N.,&Konuş,U.(2012).Bettercustomerinsight–inrealtime.HarvardBusinessReview,90(9),102–108.

Nijs,V.R.,Dekimpe,M.G.,Steenkamp,J.-B.E.M.,&Hanssens,D.M.(2001).Thecategory-demandeffectsofpricepromotions.MarketingScience,20(1),1–22.

Onishi,H.,&Manchanda,P. (2012).Marketingactivity,bloggingandsales. InternationalJournalofResearchinMarketing,2(3),221–234.

Prins,R.,&Verhoef,P.C.(2007).Marketingcommunicationdriversofadoptiontimingofanewe-serviceamongexistingcustomers.JournalofMarketing,71(2),169–183.

Risselada,H.,Verhoef,P.C.,&Bijmolt,T.H.A.(2010).Stayingpowerofchurnpredictionmodels.JournalofInteractiveMarketing,24(3),198–208.

Risselada,H.,Verhoef,P.C.,&Bijmolt,T.H.A.(2014).Dynamiceffectsofsocialinfluenceand direct marketing on the adoption of high-technology products. Journal ofMarketing,78(2),52–68.

Rust, R. T, & Verhoef, P. C. (2005). Optimizing the marketing interventions mix inintermediate-termCRM.MarketingScience,24(3),477–489.

Sathi,A.(2012).BigDataAnalytics–Disruptivetechnologiesforchangingthegame.Boise:

MCPressOnline.

Sood,A.,&Tellis,G. J. (2011).Demystifyingdisruption:Anewmodel forunderstandingandpredictingdisruptivetechnologies.MarketingScience,30(2),339–354.

Srinivasan, S., Vanhuele, M., & Pauwels, K. (2010). Mindset metrics in market responsemodels:Anintegrativeapproach.JournalofMarketingResearch,47(4),672–684.

Tirunillai,S.,&Tellis,G.J.(2012).Doeschatterreallymatter?Dynamicsofuser-generatedcontentandstockperformance.MarketingScience,31(2),198–215.

VanDoorn, J.,Lemon,K.E.,Mittal,V.,Naβ,S.,Pick,D.,Pirner,P.,Verhoef,P.C. (2010).Customer engagement behavior: Theoretical foundations and research directions.JournalofServiceResearch,13(3),253–266.

Van Nierop, E., Fok, D., & Franses, P. H. (2008). Interaction between shelf layout andmarketing effectiveness and its impact on optimizing shelf arrangements.MarketingScience,27(6),1065–1082.

Venkatesan, R., & Kumar, V. (2004). A customer lifetime value framework for customerselectionandresourceallocationstrategy.JournalofMarketing,68(4),106–215.

Verhoef,P.C.(2003).Understandingtheeffectofcustomerrelationshipmanagementeffortson customer retention and customer sharedevelopment.Journal ofMarketing, 67(4),30–45.

Verhoef, P.C., Franses, P.H., Donkers, B. (2002). Changing perceptions and changingbehaviorincustomerrelationships.MarketingLetters,13(2),121–134.

4.1Classicdataanalytics

Introduction

Marketinganalyticshavebeenaround fordecades.Especially since the1960s,marketingscientistsandmarketresearchandconsultingagencieshavedevelopedseveralmodelsandtechniques toanalyzedata. Inparticular,asearlyas the1970s,econometricmodelswerebecoming popular. Leeflang and Naert (1978) published an influential book on theapplication of econometric models in marketing, and marketing scholars such as PhilipKotler andGary Lilien, produced handbooks onmarketingmodels (e.g. Lilien,Kotler,&Moorthy, 1992;Lilien&Rangaswamy, 2006).Thesemodelsweremainlyused formarketandbrandanalytics.SeveralmodelshavebeenappliedandmarketresearcherssuchasACNielsen advocated the use of specific models to, for example, predict the effects ofpromotions on sales, for example by using the SCANPRO model (Wittink, Addona,Hawkes,&Porter, 2011; Leeflanget al., 2015).Amodel frequently cited for its enormoussuccessinbusinessisConjointAnalysis,usedtomeasurecustomerpreferencesforspecificattributes (Roberts, Kayandé, & Stremersch, 2014). By the 1990s David Shepard andAssociatessuccessfullydemonstrated theuse of analytics indirectmarketing to increaseprofits(DavidShepardAssociates,1999).Earlyin2000alimitednumberoffirmsadoptedprediction models for customer behavior. With the growth of customer relationshipmanagement (CRM), the attention on and application of individual models predictingindividual customerbehavior increased sharply (Blattberg,Kim,&Neslin, 2008;Verhoef,Spring,Hoekstra,&Leeflang,2002;Verhoef,Hoekstra,&VanderScheer,2009).

Overviewofanalytics

Inthischapterwewilldiscussaselectionofclassicdataanalytics.BasedonourdiscussioninChapter4,wewillclassifytheseanalyticsbasedontheir:

applicationarea(market,brand,customer)whethertheyaredescriptiveorpredictive;andwhethertheyareastaticordynamicmethod.

InTable4.1.1weshowourselectionofsevenclassictechniques.Weinnowayclaimthatthisisanexhaustivelist;wesolelybaseourdiscussiononwhatweobserveasimportanttechniquesinmarketingpractice.Westartwithdescriptivesasinmanycasesthesearestillvery important given that they are frequently used for management reporting in, forexample,marketing dashboards. Also, profiling is still rather descriptive in nature, as itusuallyreportsdescriptivesonspecificvariablesofdifferentbrands/customergroups.Thenextclassics involvemorecomplicatedandfrequentlymultivariateanalytical techniques,suchasclusteranalysisandregression.Wewillcontinuewithadiscussionofeachoftheseseven“classics.”

Classic1:Reporting

With reporting, analysts aim to provide management information on some relevantdescriptivestatisticsonspecificKPIs,suchasmarketshare,customersatisfaction,and/orcustomer profitability. Reporting is especially important in marketing dashboards andspecifictoolingtofacilitatethemhasbeendevelopedinbusinessintelligencesoftwaresuchas IBM Cognos and SAP Business Objects. Averages or means are probably mostfrequentlyreported(e.g.averagesatisfactionscore). It isnotourobjectiveheretodiscussall these statistics indepth, asweassume theyare ratherbasic and shouldbeknown toreadersofthisbook.Ifnot,werefertobasicstatisticsbooksandmarketingresearchbooks(seeMalhotra, 2010). However,we still observe some commonmistakeswhen reportingKPIs:

Be aware of themeasurement scale:Multiple variables aremeasured using non-metricscales—forinstancechurnmightbemeasuredbyasimpleyes/no.Averagescan thennotbecalculated. In these instancespercentages shouldbe reported (i.e.churnrate).

Table4.1.1Sevenclassicdataanalytics

Averages can be misleading: Only looking at the average can be misleading,especiallywhen there ismuch heterogeneity. Therefore,we strongly recommendnotonlylookingatoverarchingstatistics,butalsoatthedistributionofvariables.Inthisregard,descriptivessuchasthestandarddeviationcanbeuseful.Graphically,box plots can be shown. A histogram is also very useful. A histogram can, forexample, show that there are two segments of customers when considering acustomer feedback metric: haters and lovers. Inventing some numbers for thismetric (seeFigure4.1.1), anaveragemight then reporta scorearound6,and thisaverageisnotveryinformative,asapparentlythereareveryunsatisfiedcustomers(scoringaround4)andverysatisfiedcustomers(scoringaround8).Focus on extremes: To outperform competition an average performance isfrequentlynolongersufficient.Firmsneedtodelightcustomersandhaveextremelywellevaluatedbrandstosuccessfullycompete.Marketingmetricreportingshouldthereforegobeyondonlyreportingaverages,andalsoconsiderextremeresponses.Thisisspecificallyimportantforcustomerfeedbackmetrics,inwhichtop2-boxesorspecifictransformationssuchasthenetpromoterscore(NPS)canbeveryvaluable(seealsoourdiscussioninChapter2.1).Report trends: Reporting on the current status can be informative, butmanagerswillbemoreinterestedinthetrend.Isthechurnrategoingdown?Arecustomersbecomingmoresatisfied?Doesourbrandimageimprove?Doesthemarketgrow?Inthatsense,specific trendmetricssuchasgrowthratescanbevery informativedescriptors. Static descriptors then becomemore dynamic andwill raise specificquestions(seeFigure4.1.2forasalestrend).

Finally, we note that descriptive analyses can be very valuable as the first step in ananalysis.Theanalystgetsmoreknowledgeaboutthedifferentdataandtheirdevelopmentovertime.Moreover,adescriptiveanalysismayshowwhether

Figure4.1.1Differentdistributionscausingsimilaraverages

therearespecificmistakesinthedataand/orspecificoutliers(abnormalobservations).Wetherefore strongly recommend analysts to first do some descriptive analyses beforeexecutingmorecomplicatedtechniques.

Figure4.1.2Exampleoftimeseriesforsales

Classic2:Profiling

Withprofiling,analystsaimtoprovideadescriptionofbrandsandcustomersegments.Forexample, researchersmaywant to compare Apple vs. Samsung vs. Blackberry users onspecificvariables,suchasage,education,andgender.Similarly,customersegmentscanbecomparedintermsofresponsetomarketingoffers,churn,customerlifetimevalue(CLV),etc. These analysesmay help firms to improve their understanding of characteristics ofsubgroups and segments, which in turn may be useful for tailoring propositions. Incustomermanagementitcanbeusefultoidentifylook-alikesforacquisitioncampaigns.Toachieve this, crossings should be made. This just means that per brand or customersegment descriptive information is provided. For simplicity we will discuss thesetechniquesforcustomerlevelanalyses.

Customercrossings

Dependingon themeasurement scaleof thecustomercharacteristic,differentdescriptivestatistics are provided. For continuous and interval scales averages are usually reported.Using analysis of variance, an F-test can assess the existence of significant differencesbetween customer groups. For ordinal and nominal scales percentages, reflecting thepercentage occurrence of a specific characteristic in a customer group, is calculated. ToassessthepresenceofsignificantdifferencesbetweengroupsChi-squaretestscanbeused.Thesetestsarefrequentlysignificantwithlargesamples,andoneshouldbecarefulbyonlylookingatsignificancelevelshere(seealsoChapter4).Inthisbookweassumesomebasicknowledgeofthesetests,andwerefertothebasicmarketingtextbookofMalhotra(2010)foramorein-depthdiscussionforinterestedreaders.

One important issue is how to present the outcomes of customer crossings in a waywhich allows readers to quickly understand differences between customer groups. Oneuseful way is to index the value for each group, where the index is the value of thedescriptive (e.g. average)within a group divided by the average for thewhole customerbasemultipliedbyonehundred. InFigure4.1.3we show this indexation for profilingofnewcustomersonageclassification.Asakindofruleofthumbanalyststypicallyconsideranindexlargerthan110orsmallerthan90as“substantial.”However,oneshouldbecarefulwithsmallcellsizes,asintheseinstancesdifferencescanbeverylargeintermsofindices,whiletheyarenotactuallysignificant.

Decileanalysis

A frequently used technique to divide customers based on their value is decile analysis.

Withthismethodcustomersaredividedintotenequalsizedgroups,eachconsistingof10%of the customer database. These groups are ranked,where the segmentwith the lowestaveragevaluegetsrank10andthesegmentwiththehighestaveragevalueisrankedas1.Inasubsequentanalysis,rankedsegmentscanbecharacterizedbyspecificvariables.Thesevariables may include customer characteristics such as customer revenues, margins, orresponses tomarketing activities. For example, for each customer profitability decile theaverageagecanbecalculated,ortheaverageretentionrate(seeFigure4.1.4).Wecarriedoutadecileanalysisforabookclub,basedonthemonetaryvalueofeachcustomer.Inasubsequent analysis,we calculated the average retention rate in testmailing.As can beobservedfromFigure4.1.4,theaverageretentionrateishighestindecile6,whichdoesnothavethehighestmonetaryvalue.Actuallytheretentionrateisrelativelylowforthemostvaluablesegment.

Figure4.1.3Profilingnewcustomersonageclassification

Figure4.1.4Decileanalysisformonetaryvalueandretentionrates

A decile analysis is often referred to as a gains chart analysis for customer responseanalysis.Thisgainschartisactuallyverysimilartodecileanalysis.Thedecilesarerankedbasedontheirresponseprobability.Somesubsequentcalculationscanbedonetocalculatethemarginperdecileandthedecileswithapositivemargincanbeselectedfortargeting

(seeFigure4.1.5).

Externalprofiling

So far, we have mainly discussed analyses where we compare customer segments withother segments. One could label this an internal customer profiling analysis. However,managers are also interested in the profile of their customers (segments) in comparisonwiththerestofthemarket.Forthispurpose,firmsuseexternalprofilinganalyses.Inthisanalysis, characteristics of a firm’s customers (groups) are compared with customercharacteristics in themarket or population. Again, indexing is frequently used tomakethese comparisons. Using these indices, one could for example say that customers of aprivatebankaretwotimesaswealthyastheaveragebankcustomer.

Figure4.1.5Gainchartanalysisforbookclub

Zipcodeanalysis

One specific external profiling analysis is the incorporation of zip codes. External dataproviders, such as Acxiom and Experian, have specific zip-code-level information (seeChapter3).Usingthisinformationfirmscangainanunderstandingaboutwhichzipcodesandthuslocal/regionalareas,over-orunderrepresenttheircustomers.Alongwiththiszipcodeinformation,theseexternaldatasuppliersalsohavedevelopedspecificsegments,suchas“rural families”and“singlehouseholds.”Forexample,anonlineretailermayfind thatrural families are over-represented in their customer base,while the single household isalmostnot representedat all.Again indexingcanbeveryuseful.Tocalculate the index,oneneeds to divide the frequency percentage of zip code segmentswithin the customerbase/groupbythefrequencypercentageofthesezipcodesegmentswithinthepopulation.

InFigure4.1.6weshowanactualexampleofaclothingretailer.Thiscompanyaimstocompare its clientele with the population. As can be observed from the analysis, thePrestige Positions, Aspiring Homemakers, Family Basics and Transient Renters are

overrepresentedinthecustomerbase.

Similar analyses can be done at the brand level. In Figure 4.1.7 we show how eachsegment contributes to the sales of fair trade and total coffee. As one can observe, thedemandingsegmenttakescareof24%oftotalcoffeesalesand40%offairtradecoffeesales.Thisimpliesthatthedemandingsegmentisaveryinterestingsegmentforfairtradecoffeewithanabovefairshareof167[(40/24)×100].

Somepracticalguidelines

Therearesomespecificissuesananalystcanencounterwhenrunningaprofilinganalysis:

Figure4.1.6ExternalprofilingforaclothingretailerusingZipcodesegmentation

Thenumberofvariablescanbelarge.Profilingbecomesthenratherdifficult,asthenumber of comparisons on specific variables becomes too much. We thereforestronglyrecommendtoeitherfocusona limitednumberofpre-selectedvariablesor to reduce the number of profiling variables by using principal componentsanalysis(PCA),asexplainedinBox4.1.1laterinthischapter.One should be careful with over-interpreting differences between groups. Onecommonmistake is that the analysis reveals that a firm is overrepresented in aspecific lifestyle segment. Firms may then only want to target this segment.However,thissegmentcanberathersmall.Sooneshouldlookbeyondtheprofiles,butalsoconsiderfactorssuchassegmentsize.Differences in profile analyses can become easily significant, especially whenanalyzinglargedata.AsalreadydiscussedinChapter4,oneshouldnotonlyfocusonsignificancebutalsoonthesizeofthefounddifferences.

Profilingisaunivariateanalysisinwhichoneconsidersstudieswithamaximumoftwovariables.Inessence,thesearejustassociationsandoneshouldbeverycarefulin interpreting these associations, as there can be spurious correlations. Causalinferencescannotbemade!Crossingswithcontinuousvariablesarenotveryinsightfulgiventhelargenumberof possible cells. One way to overcome this is to create sub-groups in thecontinuousvariable.Actuallythedecileanalysisisawaytodothis(anditcanbedonewithfewersubgroups—e.g.quartiles—aswell.Thismayhelpanalyststogainmoreinsightsonhowavariableisassociatedwithothervariables.

Figure4.1.7Salessharepercustomersegmentfortotalcoffeeandfairtradecoffee

SourceadaptedfromGfKpanelservices1

Classic3:Migrationanalysis

Migrationanalysiscanbeusedtoinvestigatethedevelopmentofcustomersandbrandorproduct usage over time. This migration analysis is frequently required to understandchangesinaggregatesalesfiguresovertime.Changesinnumbersoftransactions,numberofcustomers,turnoveretc.areoftenreported,especiallyinfinancialreports,butmanyofthe changes are notmajor, and this tends to suggest that everything is stable.However,therearemultiplebehaviorshiddenunderthesurfaceoftheseaggregatefigures.Importantvaluedriversmaychange:newcustomersarebeingacquired,customersmaychurn,cross-selling and down-selling and product and brand switches may occur (e.g. Verhoef, VanDoorn,&Dorotić, 2007).To capture valuedevelopment over time, analystsneed to gaininsights into how customer status (e.g. churn) and behavior is changing over time.Although on a year-by-year basis changes may be limited, over a longer time periodstructuralchanges intheunderlyingvaluedrivers(e.g. loweracquisitionrates,continueddownsell),canhavedramaticeffects.InFigure4.1.8weprovideanexampleforatelecomprovider. In Figures4.1.8 to 4.1.12we provide fictive figures to illustrate the Like-4-Like(L4L)analysis.

Onayearlybasistherearechanges,buttheyarenotverydramatic.However,overfiveyears,5%ofthecustomersarelost.Thequestionis:whathappened?Afirstanalysisoftheunderlyingvaluedrivers,inwhichthebaseisdecomposedinchurnersandnewacquiredcustomers, shows that the number of acquired customers is not sufficient in size toovercomethelossofcustomersduetochurn(seeFigure4.1.9).

Figure4.1.8Fallingsubscriptionbaseforatelecomprovider

Figure4.1.9Decomposingsubscriptionbaseinacquisitionandchurn

Migrationmatrix

Aclassicalwaytoshowmigrationswithinthecustomerbaseistouseamigrationmatrix.These migration matrices show how customers change from one period (t) to the nextperiod (t+1). InFigure4.1.10,weshowhowthecustomersat timet,purchasingdifferentservicesfromatelecomfirm,moveintermsoftheirpurchasebehavior.Forexample,ofthe72,500customerspurchasingproductsA,B,andC,15,000churned,while26,000customerskept the same product bundle. However, some customers also have a down sell. In thisfigure we could also have chosen to show percentage values. Although the migrationmatrixhasitsadvantages,italsohastwodisadvantagesthatmakeithardformanagerstounderstand:

1. The large number of combinations in these migration matrices means that thematrixisdifficulttointerpret.AlthoughFigure4.1.10seemsinsightful,italreadyhas 64 combinations (8 x 8). If in these tables percentages are also added, thenumberoffigurestobeprocessedbecomesenormous.

2. UsingamigrationmatrixonlyoneKPIcanbeshown.ForexampleinFigure4.1.10weonlyshownumberofcustomers inagroup.However,onemightalso like toknowtherevenuesandtheCLV.

Figure4.1.10Migrationmatrixofcustomersofatelecomfirm

Like-4-likeanalysis

Figure4.1.11Like-4-likeanalysisforvaluedevelopmentofthecustomerbaseofaphoneoperator.

To overcome these issues, the so-called L4L analysis is frequently used, particularly inretailing,wheremanagementaimtounderstandthenetturnoverdevelopmentaccountingforclosingandopeningofanewstore.WithincustomermanagementaL4L-analysisaimsto insightfully show the customer flow accounting for different value drivers, such ascross-selling,up-selling,down-sellingetc.Further,itcancombinevolumeandvalueKPIs.In Figure 4.1.11 we show an example of the value development decomposed by valuedriver. As can be observed, strong value is created with up-selling customers, whereasvalue is lost through, for example, outflowof customers (churn) anddown-selling.Notethat inflow also has a negative value effect on the customer base, probably becauseacquisitioncostsareratherhigh.CreatingaL4Lanalysisseemsrathereasy,butitismoredifficult than frequently initially considered.We consider the following steps to create aL4L-analysisforperiodtandt+1:

1. Calculatethenumberofcustomersatperiodt2. Determinethetotalvalueofcustomersatperiodtandperiodt+1pervalue-driver.

Foroutflowvalueiszeroatt+1andforinflowvalueiszeroatt3. Determinethedifferencesbetweenthetwotimeperiodstandt+14. Determinetherelativecontributionin%ofwhateachgrouphasonthetotalvalue

ofthebase5. Determinetheweightedvalueimpactbymultiplyingthepercentagebythevalue

difference.

In Figure 4.1.12we display the steps on how to execute an L4L analysis, which clearlyshowstheexecutionofthefivestepsdiscussedabove.

Morein-depthanalyses

AnL4L-analysisprovidesinsightsintothechangesofthevalue-driversovertime.Anext

step is to execute a more in-depth analysis, an in-depth analysis per value driver. Oneexample is the so-called “cohort analysis,” in which acquired customers (inflow) arefollowed over time. In Figure 4.1.13, we show the results of such a cohort-analysis,revealingthatfirstmonthchurnhasimprovedandchurnisstabilizinginthefourthmonth.

Asimplesurvivalanalysiscanalsobedonetoobservethechurnovertimeofcustomersinspecificcustomergroups(seeFigure4.1.14).Inthissurvivalanalysis,oneobserveshowcustomer groups A, B, and C develop over time, and specifically what percentage ofcustomershaveremainedcustomersafteraspecifictimeperiod.

Moreadvancedanalytics

Figure4.1.12StepsforexecutionofanL4Lanalysis

Theanalyticsdiscussedsofarinthischapteraremainlydescriptiveinnatureanddonotrequireextensivemodeling.Thesemethodsmainlyconcernanalyzingthedatainasmartway tounderstandmigrationpatternsandsubsequentvalueconsequences inacustomerbase.Butwithinmarketing sciencemore advancedmodels are beingused tomodel andpredictmigration patterns. Specifically,Markovmodels have been used.HiddenMarkovmodels have become very popular. These models assume that there are some hiddenunobservablestatesinwhichcustomersmoveintheirrelationshipwithacompany.UsingahiddenMarkovmodel,thesestatescanbeuncoveredandcustomerscanbeclassifiedtothefoundstateineachperiod.Netzer,LattinandSrinivasan(2008)werethefirsttoapplythismethodology,whenanalyzinggift-givingbehaviorofalumniofaUS-baseduniversity(see alsoMark, Lemon, Vandenbosch, Bulla, &Maruotti, 2013). One can also use thesemodels to assess potential effects of marketing instruments (e.g. direct marketing) onmovingcustomersbetweenstates.Typically,alowactivestateisfoundforcustomerswhohave infrequent transactions with the company, while also a state for loyal, frequentlyinteractingcustomersisusuallyfound.

Classic4:Customersegmentation

Figure4.1.13Exampleofacohortanalysis

In our previous discussion on customer profiling,we explained some simplemethods tosegment the customer base. There are alsomore advancedmethods that can be used tosegmentthecustomerbase,whichusestatisticalalgorithmstofindcustomersegments.Intheaboveanalyses,databaseanalystsusuallydefinetheirownsegmentationsandusuallyonlysegmentonalimitedsetofvariables(e.g.monetaryvalueinthedecileanalysisgivenas an example). Statistical segmentation techniques allow researchers to segment onmultiplevariablesandtobasethesegmentationonstatisticalcriteria,suchasstatisticsonmodelfit(Wedel&Kamakura,2000).Researchershaveseveralmethodsattheirdisposaltoexecuteasegmentationanalysis.MethodsavailableinsoftwarepackagessuchasSPSS/IBMincludeK-Means, two-stepclusteranalysis, andhierarchical clusteranalysis.Researchersface several decisions in this analysis, ofwhich deciding the number of segments is themostimportant.Thisdecisioncanbemadeonstatisticalfit-indicesand/ormoresubjectivegrounds,suchastheinterpretationoftheseveralclustersolutions.

Figure4.1.14Exampleofasurvivalanalysis

Executionofclusteranalysis

For now, we assume that most analysts still rely on the available cluster methods instatistical packages, such as K-Means.We consider five important stepswhen executingsuchananalysis:2

1. Selectionofclustervariables2. Datapreparation3. Runningtheanalysis4. Selectingnumberofclusters5. Profilingtheclusters.

Selectionofclustervariables

Typically two general types of variables are distinguished: internal (segmentation)variables and external (profiling) variables. Segmentation variables are used for creatingsegments,whereasprofilingvariablesareused todescribe thesegments.Theselectionofsegmentationvariablesshouldideallybebasedonsomeunderlyingideaofwhichsegmentscouldbepresentand,morespecifically,onwhichbasesoneaimstosegmentthemarket.Segmentation can be done using, for example, socio-demographic variables, values,lifestyles, perceived benefits, brand and product usage, or customer profitability. If oneaimstodevelopbenefitsegmentation,thesegmentationvariablesshouldmeasurebenefitsofproducts.

Datapreparation

For cluster analysis there are two main problems. First, the number of segmentationvariables is frequently large. Second, themeasurement scale frequently does not fit.Wewill first discuss the second problem. Although cluster analysis is rather flexible, morecontinuous scales are preferred. Scales with multiple non-ordered categories createproblems.However, binary variables (e.g. gender) can, if required, be included.The firstproblemleadstomassiveinterpretationproblemsofclusters.Wethereforefrequentlyfirstuse adata-reduction technique, such asPCA, to endupwith a lowernumber of clustervariables(seeBox4.1.1).Thecomponentscorescansubsequentlybeincludedintheclusteranalysis.Note thatoneshouldbecarefulwhenusingthis technique,because thederivedprincipal components only explain limited variance in the underlying variables. Thevariablesshouldthenbeincludedintheanalysis.

Box4.1.1Shortexplanationofprincipalcomponentsanalysis

PrincipalcomponentsanalysisThemainuseofprincipalcomponentsanalysis(PCA)withintheanalyseswediscussis as a data-reduction technique. Analysts typically analyse a large number ofvariables,whichfrequentlystronglycorrelate.Althoughanalyzingthislargenumberofvariablesisinprinciplepossible,itcreatesproblems.Thenumbermayjustbecomesolargethatinterpretingthedifferentoutcomesforthedifferentvariables,becomestoocomplex.Especiallyinregression-typemodels,thelargenumberofvariablescancreate problems in the estimation due tomulti-collinearity. As a consequence, theestimatedcoefficientsandaccompaniedsignificancelevelsarenolongerreliable.Onesolution for this problem is to use PCA (also referred to as “factor analysis”). Thistechnique reduces the number of variables into a lower number of uncorrelatedprincipalcomponents(PCs)orfactors,whichrepresentanunderlyingdimension.Forexample,ifonehasvariablesonInternetusage,mobileusage,andtabletusage,thesevariables could end up in one PC, representing the digital level of customers. TheresultingPCscan,forexample,beusedtoprofileorasinputinaclusteranalysis.TheuseofPCsinaregressionmodelreducesmulti-collinearityproblems,asthePCsareuncorrelated. PCA is not solely used for data reductions: it can also be usedsubstantivelyinsurveyscaledevelopmentandbrandpositioningresearch(e.g.,Lilien&Rangaswamy,2006).ThekeychallengeistofindtherightnumberofinterpretablePCs.

ExecutionExecutionofPCArequiresfourmainsteps:

Variableselection,preparation,andanalysisSelectionofthenumberofPCsInterpretationofthePCswithrotationSavingthePCs.

VariableselectionandpreparationFirstthesetofvariablestobeanalyzedshouldbeselected.Thesevariablesshouldataminimumbeinterval-scaled.Insomecases,binaryvariablescanalsobeincluded,butthis isgenerallynotpreferred.Nominalvariablescannotbeused. Ifone focusesondata reduction, there is no other selection criterion, as one just aims to analyse alowernumberofvariables.Topreparethedatathevariablesneedtobestandardized,meaning that they get a mean of 0 and a standard deviation of 1. Most softwarepackageswill automatically execute this standardization. The analysis can then beexecutedinstandardpackages.

SelectionofthenumberofPCsOnekeychallengeistoselecttherightnumberofPCs.Themainanddefaultcriterionusedisthatoftheeigenvaluebeinglargerthanone.Thiscriterionclearlyfocuseson

datareductionastheeigenvaluemeasuresthepartoftheexplainedvarianceofaPC.If the eigen value is smaller than one, the PC explains less variance than a singlevariable.AmoresubstantivemethodistofocusontheinterpretationofPCs.

InterpretPCsThe interpretation of PCs considers the variables grouped together and theimportance of these variables. For the interpretation, rotation techniques are used.ThemostimportantandfrequentlyusedtechniqueontheinterpretabilityofthePCsis to use Varimax rotation. PCs can be difficult to interpret when the variablesgrouped into a PC do not seem related. One could then consider solutions fordifferentnumbersofPCs.TheimportanceofeachvariableinthePCcanbelearnedfrom the rotated factor loading, which varies between 0 and 1. Typically onlyvariableswithloadingslargerthan0.30or0.40areconsidered.Thelargertheloading,themoreimportantthevariableinthePC.IfvariablesdonotcontributetoanyPCs(only lowloadings), theyareexcludedfromtheanalysis.Typicallyanalystsalsodonot prefer variables to have high loadings in multiple PCs, as that createsinterpretation problems. One frequently would then try different solutions. In thetable below we show the outcome of a PCA for the perceptions of attributes ofsupermarkets.

SavingPCsIf the analyst has come to a final number of PCs and has a final solution, the PCscorescanbe saved in thedatabase. Importantly, these scoreshaveanaverageof0andastandarddeviationof1.Thescoresaredifficulttointerpret.Ahigherscoreon,forexample,aPCmeasuringdigitalcustomerbehavioronlysuggeststhatcustomersareshowingmoreofthisbehavior.However,noexactvaluescanbelinkedtothis.Todoso,transformationbacktotheoriginalvariablesisrequired.

Source:Hunnemanetal.(2015)

Runningtheanalysis

Many cluster methods are available within the standard software packages, and withinthese methods also specific options are available. An analyst should have sufficientknowledge about each of thesemethods and their options tomake an informed choice.Typically, analysts also combine specific methods. For example, the hierarchical clusteranalysisonasmallsampleisveryusefulasafirststepinananalysis(hierarchicalclusteranalysis is less able to analyze large datasets, and also the selection of clusters ismoredifficult with larger number of data). Based on this hierarchical cluster analysis, thenumberortherangeofclusterscanbedetermined.TheseclustersandtheiraveragevaluescanthenbeusedasinputintoaK-meansanalysis,whichcanhandlelargerdatasetsmoreeasily.

Selectionofnumberofclusters

Adendrogram can be used to select the number of clusters. A dendrogram shows howspecific cases are combined into clusters (see Figure 4.1.15). Based on a subjectiveassessment of the dendrogram a range of cluster solutions can be considered.3 WhenrunningthesubsequentK-meansanalysis,severalsolutionsforthisrangeofclusterscanbeachieved.Adefiniteselectionshouldthenbebasedonsegmentationcriteria,suchassizeoftheclustersandinterpretability.Thiswillprobablybedonesimultaneouslywiththenextstep,asprofilingtheclustersolutionshelpsininterpretingtheclusters.

Profilingclusters

Figure4.1.15Exampleofadendrogram

The profiling of clusters can be done on the internal cluster variables and the external(profiling) variables not included in the cluster analysis. This helps to gain a furtherunderstanding of the found segments, as profiling variables such as socio-demographiccharacteristics,mediausage,channelusage,andbrandpurchasebehaviorarebeingused.Onewaytoprofiletheclustersistousediscriminant-analysis,4whichcanbeusedtoshowwhichvariablesexplainspecificclustermemberships.Onecanalsouse this technique toclassify customers in a cluster. Based on profiling, target segments can be chosen andtarget-marketing strategies can be developed. A smart way to understand differencesbetween segments is to use a scatterplot inwhich the segments are related to themostdiscriminatingprofilevariables(seeFigure4.1.16).

Advancedclustertechniques

Moreadvanced statistical clustermethodsareavailable in specialized softwarepackages,suchasLatentGold.Amongmarketingacademics,latent-classclusteranalysisinparticularhas received considerable attention (Wedel & Kamakura, 2000). This method has beenapplied to derive customer segments in different areas, such asmulti-channel usage andfinancial service buying behavior (e.g. Konuş, Verhoef, &Neslin, 2008; Paas, Bijmolt, &Vermunt,2007).Oneimportantadvantageofthistechniqueisthatitprovidesfit-statisticson which the optimal number of segments can be more objectively determined.Furthermore, it easily allows the linking of covariates to a segment solution, making itrelatively easy to detect how segments differ in terms of background characteristics. Adetaileddiscussionofthesetechniquesgoesbeyondthescopeofthistextbook.WerefertoWedelandKamakura(2000)foramoredetaileddiscussion. InFigure4.1.17weshowthestudyofKonuşetal.(2008)onmulti-channelusageasanexampleofthistechnique.

Figure4.1.16Visualizationclusters

Figure4.1.17Exampleofaclusteranalysisofshoppers

Source:AdaptedfromKonuşetal.(2008)

Somepracticalguidelines

When executing cluster analysis in practice, we encountered some practical issues andconsiderationsrequiringattention:

The abovediscussion of doing two separate cluster analyses (hierarchical andK-Means) is considered rather complicated and time-consuming. Furthermore, thelarge sizeof currentdatasetsmeans thatK-Meansbecomesmoreuseful.Anothercommon way of running the analysis is for the analyst to execute a K-Meanscluster analysis for a number of cluster solutions (e.g. 3–6 clusters) and thencompare the different solutions and select the one that performs best oninterpretabilityandusefulness.Oneway to reduce computation time canbe to execute the cluster analysis onasmallersampleandsubsequentlyuseadiscriminantanalysistoclassifycustomersintoasegment.Tocreatemoreimpactwithclusteranalysis,westronglyrecommendreflectingonthe found clusters with marketing executives. This will be a first test for theusefulnessofthesegmentation.Although customers may principally belong in many segments (e.g. Wedel &Kamakura 2000), we recommend assigning them to one cluster to create moresimplebusinessrules.

Classic5:Trendanalysismarketandsalesforecasting

Firmsarefrequentlyinterestedinhowamarketdevelopsorhowbrandsaleswillgrow.Forthemthisisimportantsothattheycanconstantlymonitormarketattractiveness,andforplanning purposes they aim to know which level of brand sales they can realisticallyexpect.Forthispurpose,firmsmightbeinterestedintrends(e.g.isthemarketgrowing?),but also in actual forecasts. Notably, trends can be used in forecasting models as well.Althoughforecastingcanbedonewithothertechniques,suchassubjectiveforecastingandthroughconjointstudies(seeClassic6),wefocushereonmodelsthatusetimeseriesontheforecasted variable and predictor variables. Note that forecasts are not only done formarketorbrandsales,butcanalsobeusedtopredict,forexample,satisfactionscoresovertime (e.g. Gijsenberg et al., 2015). Typically linear regression models with continuousdependentvariablesareused.ThebasicsofregressionmodelsarediscussedinBox4.1.2.

Box4.1.2BasicsoflinearregressionmodelsThe linear regression model is probably the most frequently used multivariatetechnique. It seemsvery easy touse, but there aremanypitfalls that require someunderstanding. There aremultiple crucial stepswhen building amodel, wherewespecifically focuson thephaseswhere thepurposeand scopeof themodel is clearanddataareavailable,andthefocusisachievingmodelresults.Thecrucialstepsare:

1. Modelspecification2. Estimation3. Testingandvalidation

ModelspecificationThelinearregressionmodelisusuallydefinedforacontinuousdependentvariableYwhichisobservedovertimeandasetofexplanatoryvariablesXas:

Asaresearcher,onewouldbeinterestedintheestimationsoftheβparameters,therespective standard errors, and significance levels. This model assumes linearrelationshipsbetweenthedependentvariableandtheindependentvariable(s).Thisis,however,notalwaysthecase.Forexample,foradvertisingtherecouldbedecreasingreturns on increased advertising. Or satisfaction could have a non-linear effect onpurchase behavior (e.g., Van Doorn & Verhoef, 2008). To account for this, the Xvariablesshouldbetransformedusing,forexample,alog-transformation(decreasingeffect)oraquadraticeffect.Otherpossible transformationsarea square rootandareciprocal relationship. These models are then referred to as non-linear additivemodels.Theestimationisstilldoneinalinearway;oneaccountsfornon-linearitiesbytransformationoftheX-variables.Thereisasecondsetofmodels,multiplicativemodels,inwhichtherelationshipbetweenthedependentandindependentvariable(s)areallverynon-linear.Frequently,analystswillthendoalog-logtransformation(i.e.boththedependentvariableYandtheindependentvariablesXarelog-transformed).The advantage of this method is that the estimated parameters for marketingvariables,suchaspriceandadvertising,canbeinterpretedaselasticities.

EstimationThestandardlinearregressionmodelisestimatedusingordinaryleastsquares(OLS).ThebasicunderlyingideaofOLSisthattheestimationmethodaimstominimizethedifferencebetween theestimatedvalueofYand the truevalueofY (referred toasresidualvalueordisturbancetermε).Thisimpliespracticallythattheobjectiveoftheestimation is tominimize the squared sumof all thesedifferences.TheOLSmodelhas several assumptions. They mainly concern the residual value. One importantassumptionisthattheresidualsshouldbenormallydistributed.Themodelqualityisfurtherassessedbyconsideringtheexplanatorypower.ThemostimportantmeasureforthisisR2,whichcantakethevaluebetween0and1,with0havingnoexplanatorypowerand1havingperfectexplanatorypower.TheR2measureisarelativemeasure

anditsvaluedependsonhowwelltheregressionlinefitsthedataandtheamountofdispersion in the values of the dependent variable. Researcherswould like to havehigh values for R2. However, one should be careful with this. High values can beachievedwithwronglyspecifiedmodels, forexamplewhenpartoftheY-valuesareincluded in the X-variables (e.g. prices are set based on expected sales).Moreover,somedatahavesuchalargedispersionthatitisjustdifficulttoexplain.Highvaluescanalsobeachievedbyusingmoreexplanatoryvariables.Toaccountforthiseffect,theso-called“adjustedR2”canbecalculated,whichwillbe lowerthanthevalueofR2.

TestingandValidationBeforeinterpretingthemodelresults,specificmodelissuesneedtobechecked.Thisspecifically concerns the assumptions underlying the linear regression model.Importantissuesthatrequireattentionare:

Theexpectedvalueofthedisturbancetermisnon-zeroThenormalityofthedisturbancetermsTheerrortermsareproportionaltothevaluesofthedependentvariable(alsoreferredtoasheteroscedasticity)The presence ofmulti-collinearity (strong correlations between independentvariables)Thepresenceofauto-correlation,implyingthattheerrortermsovertimearecorrelated(onlyrelevantformodelswithtimeseries)The independent variable is correlated with the disturbance term (alsoreferredtoasendogeneity).

If one of these specific issues happen, specific problems such as biased, unreliableparameter estimates andnon-trustedp-values occur. In practice, strong attention isgiven to multi-collinearity. However, the other assumptions are also important.Leeflangetal.(2015)heavilyemphasizetheimportantassumptionthattheexpectedvalue of the disturbance term is non-zero. This occurswhen themodel is notwellspecified and specifically when not all relevant predictors are included. Theythereforestronglyemphasizetheimportanceofawell-specifiedmodel.Recently,wehave observed considerable attention being paid to endogeneity. This occurs, forexample,whenmanagersbasetheirusedmarketingmixontheexpectedeffectsandsaleslevels.Anothercauseforthiscouldbethatthereisself-selection.Forexample,customersbecomeamemberofaloyaltyprogrambecausetheybelievetheywillbuymoreinthefuture(e.g.,Leenheer,Bijmolt,VanHeerde&Smidts,2007).Forpredictivepurposesendogeneityseemslessofanissuethanformoredescriptivepurposesofamodel(Ebbes,Papies&VanHeerde,2011).Ifthemodeldoesnotviolatethesecriteria,the estimated parameters could be interpreted with the respective p-values. Animportantaspecthereistoconsiderthefacevalidityofthefindings.Forexample,a

positiveeffectofpriceonsalesgenerallydoesnothavefacevalidityandifitoccursthealarmbells of ananalyst should start to ring!Ultimatevalidationof themodelandspecificallyforpredictivemodelsisthepredictivequality.Werecommendtheuseof hold-out samples and plotting the predictive value and the actual value in theestimation sample and the hold-out sample.Beyond that several predictionmetricscanbecalculated:

Averagepredictionerror(APE)Mean squarederror (MSE) (with thismetric largeerrorsareweightedmoreheavilythansmallerones)Mean absolute percentage error (MAPE) (this is amore relativemeasure astheerror’stermsarecorrectedfortheactualvalue).

In thenextsections,wewilldiscussseveralversionsof theregressionmodel thatcanbeusedforprediction.Inessence,thesemodelsaresimilarintermsofhowtheyareestimated.However, the model specification is different. In our example, we focus on salespredictions,butwenote that similar approaches canbeused formanyother continuousdependentvariables.Weassumethattimeseriesdataarepresent.

Trendanalysis

Figure4.1.18Trendanalysis

With trend analysis, the main underlying idea is that sales follow a specific trend. Forexample, for a new product the sales figuresmight rise over time. A first step in trendanalysis is just to plot the sales development over time (see Figure 4.1.18). This plottingimmediatelyprovidestheanalystinformationonthepresence(orabsence)ofatrendinthedata.Itmayalsohintatspecifictrends.Ifitisalineartrend,theremightbeastraight-linedevelopment in sales. However, if it is non-linear sales might, for example, growexponentiallyovertime.Thespecificationofthemodelshouldaccountforthesenon-lineareffects, by including non-linear trend variables. The basic specification of the regressionmodelincludingalineartrendisasfollows:

Usuallyananalystwillbemainlyinterestedintheβcoefficientanditssignificancelevel.Apositiveandsignificantcoefficientimpliesapositivelineartrend,whereasanegativeandsignificantcoefficientimpliesanegativelineartrend.Asnotedthetrendcanbenon-linear,something which can deduced from a plot of sales over time. To account for this,specificationsmightincludeaquadratictrend(t2)orexponentialtrend(et)ifthesalesgrowmorestronglyovertime,orasquareroottrend(√t)oralogtrend(lnt)ifthesalesgrowless strongly over time. We strongly recommend analysts to consider different trendoperationalizations.

Animportantissueisthattrendsshouldnotbeinterchangedwithseasonalitypatterns.With seasonal patterns sales follow a specific shape during the year, because demanddepends on specific developments during the year (e.g. beer sales might go up duringsummer).Toaccountfortheseeffectsseasonaldummyvariables(variablessetatvalue1ifa specific season occurs) should be included in themodels. These variables pick up thespecificseasonaleffect.

Regressionmodels

Oneproblemwithtrendanalysisisthatitonlyconsiderstheeffectsoftimeonsales,whileafirm’smarketingofcoursealsoinfluencesmarketingmetricssuchassales.Typicallythetrendmodels will have a relatively low explanatory power. This will especially hold instable markets with existing products and brands. For new products and/orgrowing/declining markets trend variables may have a reasonable explanatory power.Modelshavebeendevelopedtoassesstheeffectsofspecificinstrumentsandtopredicttheimpact of changes in marketing instruments on sales. We note that these models aretypicallynotusedtopredictsalesassuchovertime.Researchershavebeenmoreinterestedinwhathappenswithsalesif,forexample,theadvertisingbudgetisincreasedby10%.Inaratherbasicformthesemodelsareformulatedasaregressionmodelinwhichsalesdependonthevalueofthe includedinstruments. Inthesemodelsonecanalsoaccountfortrendandseasonaleffectsifrequired.Thetypicalformofthisregressionmodelincludingthreemarketinginstrumentsandatrendisasfollows:

Thismodelcanbeformulatedinsuchawaythatbothsalesandthemarketinginstrumentsarelog-transformed,butfornowweassumelineareffectswithnon-transformedvariables.Generally,onewouldexpectpositiveβcoefficientsforadvertisinganddistributionandanegative β coefficient for price. One important problem is that the true effect of thesevariablesisnottheestimatedeffect,becauseoftheendogeneityproblemarisingfromthe

fact that managers may change their marketing based on expected sales. Typically, theestimatedeffect is larger than the trueeffect.Thechallenge is,however, to find the trueeffect. Analysts should be aware of these problems. One solution is to correct for this.Econometricians have proposed several methods, including the use of so-called“instrumental variables” (see Leeflang, Wieringa, Bijmolt and Pauwels (2015) fordiscussion).Aproblemarising is that sometimes the remedy isworse than the problem,especiallyifbadinstrumentsareused(Rossi,2014).Anotherpracticalsolutionistoaccountforendogeneitybyassumingsomewhatlowercoefficients,basedonexperiencecombinedwiththeestimatedfindings.

Amorecomplicatedmodelspecificationisneededtoaccountfordynamiceffects.Thesedynamic effects are required in order to take into account potential delayed responseeffectsonmarketinginstruments.Forexample,advertisingint-1canstillhaveaneffectonsalesattimet.Initssimplestformthismodelisspecifiedasfollows:

wheres is thenumberof timeperiodsbetween the timeamarketing instrument isused(i.e. advertising spending) and the sales that result from the use of that marketinginstrument. In Figure 4.1.19 we show an example of a model explaining sales for achocolate brand (Leeflang et al. 2015). In this example own price has a strong expectednegativecoefficient,whereasdifferent typesofpromotions for thatbrandgenerallyhavepositiveeffects.Therearealsosomeseasonaleffects,asinSpringothertypesofchocolate(e.g. Easter eggs) are purchased. There are many more complicated versions of thesemodels,suchasthegeometriclagmodelandtheKoyckmodel.WerefertoLeeflangetal.(2015:Chapter2)foradetaileddiscussion.

Timeseriesmodels

Time seriesmodels are a specific form of regressionmodels. In theirmost simple formthesemodelsincludethelagofadependentvariableasanexplanatoryvariable.However,thisisnotnecessary.Ifsalestimeseriesfollowaverystablepatternandthevaluesseemtovaryaroundanaveragemodel,amovingaverageprocesscanbemodeledinwhichsalesattimetsolelydependondisturbancetermsatt–s.

Figure4.1.19Effectsofdifferentmarketinginstrumentsonsalesforachocolatebrand

Source:AdaptedfromLeeflangetal.(2015:113)

Inmanycases,however, timeseriesdonot followa stablepattern. In thesecases it iscommon to include trends, lagged variables, and also explanatory variables. The laggedvariable is sometimes also referred to as an autoregressive term and models can havemultiple lags of the sales variable in the model. Typically, models with lagged salesvariableshaveastrongexplanatorypowerandtheirpredictiveabilitycanalsobestrong.Initsmostsimplespecificationalinearmodelwithlaggedsalesandmarketinginstrumentsisspecifiedasfollows:

By including this lag variable of sales a regression model accounts for the fact thatmarketing instruments can have a long-term effect. Time series models seem relativelyeasytograsp.However,therearemanypitfalls,notleastbecausetimeseriescanbenon-stationary.Astationaryprocessisastochasticprocesswhosejointprobabilitydistributiondoes not changewhen shifted in time. Consequently, parameters such as themean andvariance,iftheyarepresent,alsodonotchangeovertimeanddonotfollowanytrends.Ifthe time series parameters in the model are non-stationary then any results cannot beinterpreted,andresearchersshouldtransformthedatatomakethemstationary.Ifatrendiscausingtnon-stationaryseries,thetime-seriesdatashouldbede-trended.Awaytodothis is to include a trend variable in themodel. If the data are then stationary they areconsidered as trend-stationary. However, frequently this is not sufficient. A frequentlyapplied method is then to take first differences. This implies that the non-stationaryvariablesaretransformedasfollows(examplesales):

If thedataare stationaryafter first-differencing theyare said tobedifference-stationary.

The data can then be analyzed. It is thus crucial when studying time series data thatresearchersassessthestationarityofthedata:severaltestsareavailabletodothis(e.g.theaugmentedDicky-Fullertest).Westronglywarnagainstjustestimatingallkindsoftimesserieswithoutasolidunderstandingoftimeseriesanalysis,astherearemanypitfalls.5

Time-seriesmodelshavebeenextendedandspecificallysowiththeintroductionofso-calledvector-autoregressive(VAR)models.TheseVARmodelsarenowfrequentlyusedinmarketingsciencetoestimateshort-and long-termeffectsofmarketing instruments (e.g.Pauwels,2004;Dekimpe,&Hanssens,1995;Nijs,Dekimpe,Steenkamp,&Hanssens,2001).In thesemodels a system of equations is formulated in such a way that each includedvariableisinfluencedbythelagsofotherincludedvariables.Soinitssimplestform,withsales and two marketing instruments, price and advertising, there are three equationsestimatedsimultaneouslyinwhichonealsoallowsthedisturbancetermstobecorrelated.In a rather simplistic form this model can be specified as follows, assuming stationaryseries:

Theabovemodelcanbeextendedbyincludingsomecontrolvariables(X),turningitintoaVARXmodel.Severalothervariantsofthesemodelshavebeendeveloped.Tounderstandthe short- and long-term effect, impulse response functions have to be estimated. VARmodelsgaininpopularityasmoretimeseriesdatabecomeavailable.

In general it is believed that time seriesmodelswill gain in popularity.Dekimpe andHanssens (2010) specifically consider the expanding amount of available data, theincreasing dynamics in the marketing environment, the greater need for marketingaccountability, and the emergence of online data, as important drivers. They specificallystate that theyareconfident that the importanceof timeseriesmodels inmarketingwillcontinue togrow (Dekimpe&Hanssens, 2010: 27).Despite thisdevelopmentweobservethat marketers are still reluctant to use these models, whereas they have become verycommon in, for example, finance and economics (i.e. for forecasting economic growth).OnemainproblemwithVARmodelsisthattheycouldeasilybecomeakindofblackboxinwhich everyvariable is influencedbymanyothervariables, and theanalystmaynothaveasufficientunderstandingofwhatisactuallyoccurring.Wehavealreadymentionedthemanypitfallsinusingtime-seriesmodels,buttrainingintheiruseissofarlimitedtobusinessschoolsandmarketingprograms.InFigure4.1.20weshowthepredictedvaluesofaVAR typemodel (DASVAR) explainingperceived service quality (PSQ) for aEuropeanrailwaycompany.WeusedanestimationsampletoestimateourmodelandpredictedthePSQ(predictedPSQasymmetricLagSVAR ingraph) forboth theestimationsampleandthehold-outsampleThemodelpredictsthesatisfactionseriesprettywell,bothin-sample

andout-of-sample.

Practicalconsiderations

Figure4.1.20PredictionsforservicequalitytimeseriesofaEuropeanpublictransportfirm

Source:AdaptedfromGijsenberg,VanHeerde,&Verhoef(2015)

In general there are some practical considerations to take into account when usingregressionmodels:

When estimating models researchers should carefully investigate the differentassumptions(seeBox4.1.2).Westronglyrecommendcarryingoutrobustnesschecksinwhichdifferentversionsofmodelsareestimatedtoassesswhethermodelparametersremainrathersimilarornot.Iftheyremainsimilarthemodelisusuallyconsideredrobust.Do not over-value the explanatory power of models, and do have realisticexpectations. A time seriesmodelwill typically have high R2, but thismay alsodependonthedata.Regressionmodelsusingcross-sectionstypicallyhavelowerR2.However, when using survey data R2 might be higher due to common methodvarianceissues.We strongly recommend the plotting of data to identify patterns in the data andpossible(non)-linearrelationships.A frequent question concerns the sample size required to estimate regressionmodels. A general rule of thumb we frequently use is that per independentvariables around 5 observations should be present (Leeflang et al., 2015).Havingsufficient data points is important, as results may easily become instable andunreliable.Be aware of the fact that when analyzing aggregated data, the analyst ignoresunderlyingsegmentsorbrandsorcustomers.Henceparameterestimatescouldbedifferentfordifferentsegmentsorbrands.Thequestionhereiswhetheroneshouldpoolthedataoverthesesegmentsorbrands,orestimatepersegmentorperbrand,

called“unpooledanalysis”(seeLeeflangetal.,2015).Ifdatacannotbepooled,theaggregatedpooledanalysiswillleadtowrongestimates.

Classic6:Attributeimportanceanalysis

Especiallyinnewproductdevelopment,firmsaimtounderstandtheimportanceofspecificattributes.Foradigitalcameraonewouldliketoknowtheimportanceofattributessuchasnumberofpixels, size,andprice.Forhotels thedestination, thesizeofrooms,additionalservices etc. could be important attributes. Attributes are not only important for newproductdevelopment,butalsoforcreatingsatisfiedcustomers.Firmswouldliketoknowthe contribution of each service attribute towards customer satisfaction. For example,hotelswouldliketoknowhowanimprovementintheserviceatthereceptiondeskcouldimprovesatisfaction.Typically,wedistinguishbetweenthreemethodsthatcanbeusedtoassesstheimportanceofattributes:

1. Statedimportancemeasurement2. Regression-basedapproach3. Conjointanalysis.

Statedimportancemeasurement

Arathersimpleapproachistoaskcustomerswhichattributestheybelieveareimportant.Typicallycustomersareaskedtoratetheimportanceofspecificattributeson,forexample,a 5-point scale (1 = absolutely not important attribute, 5 = very important attribute).Analystswillreporttheaveragescores,aswellastop-2scores.Ofcourseahigheraveragescoremeansahigherimportance.Oneproblemwiththistypeofmethodisthatcustomersdo not make a trade-off between attributes. Frequently, customers state that manyattributesareveryimportantandtherearenotmanydifferencesbetweenattributes.Onesolutionistousedifferentsurveyquestions,suchasrankingofalistofattributes,wherethemostimportantattributeisrankedasfirst.Rankinginducescustomerstoatleastmakesometrade-offs.Notethatinthisprocessdescriptivestatisticsarebeingused.

Regression-basedapproach

Theregression-basedapproachismainlyusedinserviceresearch.Thecommonapproachistomeasureoverallcustomersatisfaction(oranotherfeedbackmetric,suchasNPS)andto link this satisfaction score to evaluationson service attributes (e.g. ona 7or 10pointscale). Usually these evaluations and the satisfaction score are measured in the samesurvey.AsastepinbetweenaPCAcanbeexecutedonthescoresoneachoftheseserviceattributes to reduce the number of attributes in the regression analysis and solvemulti-collinearity problems. Running the regression analysis, one can then observe theimportanceofeachoftheattributesbylookingattheestimatedstandardizedcoefficients.

Alargerstandardizedcoefficientimpliesthattheattributeismoreimportant.Hunneman,Verhoef and Sloot (2015) show that for Dutch retailers, the service attributes are mostimportantforcreatingcustomersatisfaction(seeFigure4.1.21forthemaineffectsintheirmodel).

Figure4.1.21Effectsofstoreattributesonstoresatisfaction

Source:AdaptedfromHunnemanetal(2015)

Oneproblemwiththeregression-basedapproachisthattheusedsurveymethodinducesastrongcorrelationbetweentheindependentanddependentvariables,whichisknownas“common-method variance.” Part of the explained variance in the dependent variable isthenduetothefactthatthedependentvariablesandindependentvariablearemeasuredinthe same survey, frequentlyusing a similar attitude scale (Podaskoff,MacKenzie, Lee,&Podaskoff, 2003; Lindell&Whitney, 2001).As a result the estimated coefficients are toohigh.Theexplainedvarianceisalsotypicallyhigh,withvaluesofforexamplearound0.40or0.50.Probablytheimportanceofdifferentattributescanbegatheredfromthesedata,ifoneassumesthatthisbiasaffectsallvariablesandtheirrespectivecoefficientsinmuchthesameway.However,using theseestimatedcoefficients topredicthowchanges inserviceattributeevaluationsaffectssatisfactionshouldbedonewithgreatcare.Onesolutionistouseanumberofsurveysinwhichtheevaluationsandthesatisfactionscorearemeasuredseparately.Anothersolutionistomeasureindependentvariablesmoreobjectively—askingnotforanevaluationbutratherwhetheraspecificeventhasbeenexperienced.

Conjointanalysis

Conjointanalysiswasdevelopedinthe1970sandisconsideredtobeoneofthetechniquesthathassuccessfullytransferredfrommarketingscienceintopractice(Robertsetal.,2014).Specific software packages have been developed by, for example, Sawtooth Software toexecute conjoint analysis, but it is also available as a standard technique in standardpackages. In essence the conjoint method involves an experimental design in whichpreferencesforspecificproductsaremeasuredandtheresultingdataareanalyzedwitha

regression or logitmodel. The conjoint experiment is awithin-subjects design inwhichrespondents are exposed to a set of artificial constructed products that vary in terms ofspecificattributelevels.Thebasicideaofconjointanalysisisthateachproductorserviceconsistsofabundleofjattributesthatjointlyresultsinaproductutility:

For each artificial product a utility or preference is measured using several kinds ofmethods.Inconjointanalysis,amodeldependentonthepreferencemeasurementmethodis estimated, aiming to assess the importance of specific attributes. One of the keyadvantagesofconjointanalysis is that respondents see the fullproductand thushave totrade off specific attributes (e.g. a low price for airline tickets vs. more leg room). Theconjointmodelcanbeusedforspecificinformation-basedproductsorservicesinwhich—based on derived individual customer preferences—recommendations are given tocustomers. The conjoint model then functions as an algorithm underlying a specificrecommendationwebsiteorapp.Thetypicalstepsinconjointanalysisarethefollowing:

1. Studydesign:Choiceofattributesandlevels2. Choiceofconjointdesign3. Collectingthedata4. Analyzingthedataandinterpretationofresults.

Studydesign

Figure4.1.22Attributeschosenforstudyoncabservices

Thefirststepinsettingupastudyisprobablythemostimportant.Forthestudiedproduct,therelevantattributesandlevelshouldbechosen.ForchoosingaflightfromParistoNewYork, the following attributes and levels could be important: airline (Air France,United,Delta);legroom(small,large);price($750;$900);andindirectflight(viaLondonHeathrow)ordirect.InFigure4.1.22weshowtheselectionofattributesforaconjointstudyinwhich

customers should evaluate different cab services.6 There are some specific issues toconsider when choosing the attributes and the accompanying levels (Eggers & Sattler,2011):

It is generally advised to keep the number of attributes low, to decrease thecomplexityforrespondentswhenevaluatingtherespectiveproducts.Sixattributesareconsideredasakindofmaximum.Ifotherattributesarerelevant,buttheyare,forexample, standardavailable forallproducts, it isadvised to include thembutnottovarythemintheexperimentalstudy.Thechosen levelsshouldberealistic.Forexample,suggestingavery lowpriceof$250foradirectflightfromParistoNewYorkisnotveryrealistic.Thenumberof levelsperattributeshouldbe limited.More levelsrequirea largerexperimentandlargersamples.AsageneralruleofthumbEggersandSattler(2011)advise not to usemore than 7 levels. In general fewer levels withmore reliableestimatesispreferredoverlesspreciseestimatesofattributeswithmorelevels.Onespecificissueconcernsthe“numberoflevelseffect.”Thiseffectariseswhenthenumberoflevelsisnotequallydistributedovertheattributes.Attributeswithmorelevelsthantheotherattributes(e.g.5forpriceand2forlegroom)leadstoahigherrelative importance of an attribute with more levels. Hence, the parameters fortheseattributesarebiasedupwards.Thistypicallyhappenswiththepriceattribute.Asaconsequencethepriceattributeisconsideredastooimportant.Theadvicetotrytoavoidthiseffectistokeepthenumberoflevelssimilaracrossattributes.Asecondimportantbiascouldariseduetothe“rangeeffect.”Iftherangebetweentwolevelsistoolargetheeffectoftheattributebecomestoostrong.Forexample,ifapricerangeof$500to$900isusedinsteadofarangefrom$750to$900theeffectofpricewillbelargerforthefirstrange.Itisgenerallyadvisedtousearangethatisrealisticwithinthemarketwhichisnottoolarge.

Choiceofconjointdesign

Thechoiceofthedesignfirstconcernshowpreferenceswillbemeasured.Therearethreemethodstomeasurepreferences:

Rating-basedconjointRanking-basedconjointChoice-basedconjoint

Withrating-basedconjointanalysistherespondentsareaskedtoevaluateeachofseveralproductswithdifferentattributesonaratingscale(e.g.0=veryunattractiveto10=veryattractive).Thesetasksarerathersimpleforrespondents.Thedisadvantageofthismethodisthatthereisnoexplicittrade-offbetweendifferentproductoptions.Withranking-based

conjoint analysis the respondent is shown several product options and is asked to rankthem from very unattractive (lowest rank) to very attractive (highest rank). With thismethodatrade-offisbeingmade,butthedisadvantageisthatthistaskcanonlyreasonablybe donewith a low number of product options, thereby creating only a low number ofattributes and levels. It is virtually impossible to rank 20 options. With choice-basedconjointtherespondentisrepeatedlyshownanumberofproductoptions(e.g.3)andthenhastochoosethemostattractiveproductfromthechoiceset(seeFigure4.1.23forchoicedesignofacabstudy).Choice-basedconjointanalysishasbecomeratherpopularamongmarketingscientists;theybelieveittobeeffective,becausechoicesareanintegralpartofpeople’s everyday life (Eggers & Sattler, 2011). Furthermore, rating scales in particularbecomeproblematicinmulti-culturalstudies,asrespondentsindifferentculturesmayusethesescalesdifferently.

Figure4.1.23Exampleofachoice-basedconjointdesignforacabstudy

Iftheconjointmethodischosentheresearchdesignthenhastobesetup.Onespecificissue is that especiallywithmanyattributes and levels, thenumberof possibleproductsincreasesdramatically.Itwouldbeveryunreasonabletoaskrespondentstoevaluateeachpotential product. Hence, more efficient designs should be developed that showrespondentsamuchlowernumberofproducts,butwhichisstillisabletoproducereliableparameterestimates.EggersandSattler (2011) refer to four importantcriteria forchoice-basedconjointstudies:

Balance:Allattributelevelsappearanequalnumberoftimes.Orthogonality:Eachattributelevelpairappearsanequalnumberoftimes.Minimaloverlap:Thealternativeswithinachoicesetaremaximallydifferentfromoneanother,i.e.avoidequallevelswithinoneattribute.Utility balance: The alternativeswithin a choice set are equally attractive to therespondent,sothatchoicesetsavoiddominatingordominatedalternatives.

Collectingthedata

Dataarenowfrequentlycollectedusingonline surveys inwhich respondentsare shownrealisticproducts.Therearesomeconsiderationstotakeintoaccountwhencollectingthedata. First, the sample size should be sufficiently large. Larger sample sizes are requiredwhenmore attributes and levels are tested. Samples should also be representative of themarket. Second, in this within-subject design respondents can become fatigued whenconfronted with multiple choice options. Researchers should keep respondents engagedduringtheconjointtask.Onewaytodosoistogivetherespondentanincentive-alignedprocedure.Forexample,theycangetafictivebudgetwhichtheycanusetospendduringthechoicetask(e.g.forspendingforaspecificalternative).Alotterycanbeusedtorewardrespondents.Onespecificissueforchoice-basedconjointanalysisisthatresearcherscouldchoosetoincludethenon-choiceoption.Thisismorerealistic,asinreallifecustomerscanalsochoosenottobuyaproductiftheirpreferredoptionisnotavailable.

Analyzingthedataandinterpretation

Thedataareanalyzedusingregression-basedtechniques.Thetypeofanalysisdependsonthechosendesign.Forrating-basedconjointanalysisregressionmodelsareused,whileforchoice-basedconjointanalysislogitmodelsareused(seethesectionsonClassics5and7).Inprinciple,amodelcanbeestimatedforeachrespondent.Hence,foreachrespondenttheimportance of each attribute can be estimated. The average values of these individualparameterswouldthenbetheaverageimportanceofeachattribute.Conjointestimationin,forexample,theSPSSsoftwaresavestheindividualestimatesandalsoshowstheaggregateresults. However, analysts can estimate aggregate models themselves. As multipleobservations per respondent are available, panel models should be used (e.g. Wuyts,Verhoef,&Prins,2009).Toimprovethepredictiveabilityofthemodels,interactioneffectsbetween attributes can also be added. For example, pricemay reduce the importance ofbrands.Importantly,thepredictiveabilitycanbetestedwithahold-outproductthatisnotincludedintheanalysis.Interactioneffectsmayindeedoccur.

Conjoint analysis can also be combined with cluster analysis to derive a benefitsegmentation.Themost simpleway todo this is touse theestimated individualutilitiessavedintheanalyticaldatabase.Theseutilitiescanbeusedasinputintoaclusteranalysis(see the section onClassic 4).Anotherway is to usemore advancedmethods inwhichlatentclassanalysisisusedtoderivesegmentswithspecificutilitiesperattribute.Forthecab study a latent-class studywas executed (see Figure 4.1.24). The resulting classes orsegments of this analysis can be interpreted as follows. Segment 1 seems to be a fan ofTeslaandtendstoslightlyprefercityservicesoverotherservices.Segment2,however,isastrongbeliever in theUberandToyotabrands,andvaluesa lowprice.Segment3valuescityservicesandprefersGoogleovertheotherbrands.Theyputlessvalueondrivingtime

thantheothersegments.

Figure4.1.24Segmentationanalysisforconjointstudyoncabservices

Advancedmethodsandmarketforecasting

A relatively recentmethod that is gaining in popularity because of the increased use ofcomputer-aidedsurveysisadaptiveconjointanalysis.Inadaptiveconjointanalysissurveyscanreactdirectlytotheanswersgivenbyeachrespondent.Asaconsequence,unimportantorunacceptable features canbe excludedduring the survey.Themain advantage is thatmoreattributescanbeused inthiskindofmethod.Moreover, itmakestheconjoint taskmoreengagingforrespondents.

Onefinalissueisthatconjointstudiescanalsobeusedformarketforecastingpurposesofdifferentalternativeproducts.Basedonthechosenproductattributesdifferentscenarioson the percentage of potential customers interested in such a product can be calculated.Morespecifically,preferencesharescanbecalculated.However,thesepreferencesharesarenot similar tomarket shares,given the fact that theconjoint studydoesnot fullymimicmarketconditionsduetoitsexperimentalnature.It isthereforestronglyadvisedtoworkwith different scenarios thatmimic themarket situation and do simulations with thesescenarios.Thiswillprovidetheanalystwithapotentialsetofpreferenceshares,thatmightbe closer toactualmarket shares. Still, one shouldbevery careful, as competitorsmightreact,customerpreferencesmaychangeetc.

Practicalconsiderations

Intheabovediscussionwehavealreadybeenratherelaborate.Despitethat,therearesomeadditionalpracticaltipswhenexecutingattributeimportancestudies:

When includingattributes in thestudy,oneshouldpreferably takeattributes intoaccountthatmarketingcaninfluence.We strongly advise not to focus solely on the aggregate findings, in which theattribute importance for the total sample is calculated. In fact there is sufficientheterogeneitybetweencustomersandcustomersegmentsonattributepreferences.Hence,amoreindividualorsegment-levelapproachiswarranted.Simulationstudiesusingtheoutcomesoftheconjointstudytopredictthepotentialfuturemarketsharesofnewproductscanbeveryusefultocreatemoreimpact.

Classic7:Individualpredictionmodels

Oneimportanttaskforcustomeranalystsistopredictfutureindividualcustomerbehavior.Forthatpurpose,analystshaveseveralmethodsandstatisticaltechniquesattheirdisposal.A naive way of predicting would be to assume that behavior does not differ betweencustomers and one assumes that the average behavior (e.g. average response rate on amailing) is the same for all customers. This is, however, not a reasonable assumption.Making it a bit more complicated one could assume that behavior differs betweensegments.Forthatpurpose,customercrossingswithbehavioraloutcomescouldbemade.Ifonethenknowsthesegmentmembership(e.g.decile1),wecanpredictfuturebehavior(seeClassic 2 on profiling and specifically decile analysis/ gain chart analysis). A specificmethod assuming this is the recency, frequency and monetary value (RFM) method.Finally,onecanusedifferentstatisticalmodelstopredictcustomerbehavior.WewillfirstdiscussRFMandthenmoveontostatisticalmethodssuchasdecisiontreesandlogit.

RFMmodel

RFMstands for recency, frequencyandmonetaryvalue.Recencyrefers to the timesincethe last transaction.Frequencyconcerns thenumberof transactions inaconsidered timeperiodandmonetaryreferstotheaverageortotalmonetaryvalueofalltransactionsinaspecific time period. The RFM model originates from direct marketing and has beenfrequentlyappliedbyfirms(Verhoefetal.,2002)suchasReadersDigest.Itisconsideredavery powerfulmodel for predicting responses tomailings andmay also be used to gaininsightsontheexpectedlifetimevalueofcustomers(e.g.Fader,Hardie,&Lee,2005).

To executeRFManalysis, analysts should calculate anRFM-index.A standardwayofformingtheRFMindexistoprovideimportanceweightsaicountinguptooneforeachoftheRFMcomponents.RFMisthencalculatedasfollows:

wherea1,a2,anda3≥0,anda1+a2+a3=1

In this formula, the standardized (meanvalue= 0, standarddeviation=1)R, F andMscoresareusedtoaccountforscaledifferences.Theweightscanbebasedonintuitionandexperience. For example, marketers might know that frequency is the most importantdeterminantofresponsetoamailing,whilemonetaryvalueislessimportant.Basedonthisinsight,monetaryvalue isassigneda lower importanceweightthanfrequency.Asecondwayofcalculating the importanceweights is to formallyestimateamodel inwhichoneestimates the effect of each of R, F andM on pastmail response or response on a testmailing.Fromtheresultsonecangaindetailedinsightsontheimportanceofeachofthesevariables.

AnotherwayofcalculatingtheRFM-indexistheexecutionofPCA.PCAisfrequentlyused to calculate indices (e.g. price index). TheRFMvariables are included in the PCA,whereonecanchoose theoptionofa singleprincipal component (or factor). If apropersolutioncanberetrieved,thePCAscorescanbesaved,reflectingtheRFMindex(seeBox4.1.1).

TheRFM index canbeused as a segmentationvariable and,using for exampledecileanalysis,customerscanbedividedintodifferentRFMsegments.ForeachRFMsegmenttheaverage response rate etc. can be calculated. Figure 4.1.25 shows how we divided thecustomerbaseof a book club into five segments for the separateRFMvariables and theformedRFMindex.ThelatterwasbasedonaPCA.TheRFMsegmentationisclearlylinkedtoresponse.ThesegmentwiththehighestRFMhasthehighestresponserate.Segmentingonrecencyonlydoesnotresultinstrongdifferencesinresponserates.Thisindicatesthatrecencyisnotagoodpredictorofresponse.

Decisiontrees

Decisiontreescanbeused inabroadrangeofstatistical techniques.Basedonthefoundassociationsbetweenoneindependentbinaryvariable(e.g.response)andusuallymultiplepredictors,atreestructurewithdifferentcustomersegmentscanbedrawn.7Inastatisticalsense one could argue that this technique involves both segmentation and prediction.However, incustomervaluemanagement (CVM)practice it ismainlyused forpredictivepurposes.SeveralnamesforthesetechniquesarefoundinliteratureandsoftwareprogramssuchasCHAID,Cart,andAnswerTreecanbeused.

Thebigadvantageofusingadecision tree is that it isquiteauser-friendly technique.Computer programs have developed user-friendly interfaces.Moreover, the outcomes oftheanalysisareeasytoexplaintomanagersgraphically.Inpractice,outcomesofdecisiontree analyses on customer churn are often presented to boardroom members. Anotheradvantageisthatonegetsmoreinsightsintodifferentrelationshipsinthedata.Especially,onegetsinsightsintopotentialinteractioneffectsandnon-lineareffects.Forexample,onemayobservethatinespeciallylargehouseholdslivinginurbanareaschurnorresponseishigher.Onedisadvantageisthatthetechniquecanbeconsideredsomethingofablackbox.Itisnotcleartoanalystswhatisactuallyhappening.Inanextremesensetheanalystputsvariables in the program and the decision tree pops out. Still, its usefulness is generallyhighlyvalued.SomeresearchersuseCHAIDasamethodtogetmoreinsightsintothedataand the interdependencies betweenvariables.Theoutcomes ofCHAIDare subsequentlyusedforfinetuningotheranalyticalmodels,suchasregression-basedmodels(e.g.Neslin,Gupta,Kamakura,Lu,&Mason,2006).

Figure4.1.25ResponseratefordifferentRFM-segments

Figure4.1.26showstheoutcomeofaCHAIDanalysisfortheresponsetoashopmailing.Theaverageresponseis1.16%.Inasubsequentnodeofthetree,householdsizeisthemostdiscriminating variable.Household sizes from2 to 5 have thehighest response.There isstill some heterogeneity in response within this segment. Therefore, a second node isadded,with distance to the shop as a discriminating variable. Especially for householdslivingclosetothestore,responseisevenhigher,withanaverageresponserateof1.89%.

Logitmodels

Withincustomermanagementandbrandingresearch,analystsfrequentlyhavetopredictbinary events such as churn or brand choice. Analysts sometimes use well-knownregressionmodels topredict thiskindofevent.Theordinary regressionmodel,however,assumesthat thedependentvariable iscontinuousandtherefore it isnotperfectlysuitedfor the prediction of these binary events. For this reason logistic regression has beendeveloped.Inthisregression-typemodelthedependentvariableisbinary.Inliteraturethemodelisoftenreferredtoasthelogitmodel.Thistypeofmodelhasbeenextensivelyusedinthebrandloyaltyliterature.Logisticregressionanddecisiontreesarethemodelsmostcommonlyusedtopredictchurn;togethertheyaccountedfor68%oftheentriesofachurnmodeling contest8 inwhich both practitioners and academics participated (Neslin et al.,2006).Neslinet al. (2006) concluded that logistic regressionand treeapproachesperformbest in terms of prediction and that the differences between the models in predictiveaccuracyare“manageriallymeaningful.”

Figure4.1.26ExampleofadecisiontreeusingCHAID

Theobjectivesofthelogisticregressionmodelaretwofold.First,oneaimstoassesstheimpact of predictors on the occurrence of an event. Second, one aims to predict theprobability of occurrence of the event. This probability p varies between 0 and 1 (or,multiplied by 100, between 0% and 100%). In the logistic regression model the binarydependent variable Y (e.g. churn) is econometrically transformed in a latent continuousvariableY*.Inourexampleweconsideredresponsetothemailingofastoreandincludedvariablessuchasdistancetothestore(DISTANCE),andhouseholdsize.Ifweworkwiththesetwovariableswegetthefollowingequation:

This latent continuous variable can then be explained by the independent variables. Foreachof thesevariables,coefficientsandaccompanyingsignificant levelsarederived.Thecoefficient(β)showstheimpactofthatvariableonthelatentcontinuousvariableY*.Thesecoefficients should be interpreted in a different way as in the regression model. Forexample,assumethatacoefficientis0.5householdsize.Ariseof1inhouseholdsizedoesimply that the latent variable Y* increases by 0.5. However, it does not imply that theprobabilityof theoccurrenceofaneventrisesby0.5.Toassess thisprobabilitya logistictransformationshouldbeused,whichcanbeformulatedasfollows:

This probability can subsequently be used in all kinds of calculations. In this case theexpectedrevenue/marginpercustomerofamailingcanbecalculated.Ifchurnisexplainedone can assess the lifetime value of a customer as one knows the averagemargin on acustomerandthechurnprobability(e.g.Donkers,Verhoef,&DeJong,2007).Thequalityofthemodelcanbeevaluated intermsof itsexplanatorypowerandthehitrate.ThemostcommonlyusedstatisticistheChi-squaretest,whichassesseswhetherthepredictorshave

some explanatory power. Programs also frequently report R2 types of measures, whichshowhowgood themodel is at explaining the event.HighR2 are frequently consideredbetter.One cautionarynote is that the statistics for the logistic regressionhavedifferentmeaningsthantheydointhestandardregressionmodel.Moreover,inthistypeofmodel,and especially in aCVMcontext, thesemeasures tend tobe low. Finally, one frequentlyuses the hit rate, which is calculated as the percentage of events predicted in the rightgroup.Oneshouldalsobecarefulwithevaluatingthisstatistic,especially ifeventsoccurinfrequently.Allcasestendtobeclassifiedautomaticallyinthemostfrequentlyoccurringevent group. Formore detailed information on the logistic regressionmodelwe refer toFransesandPaap(2001)andBlattbergetal.(2008).

Fortunately, statistical programs such as SPSS provide standard procedures to executelogisticregressionprocedures.Theseprogramsalsocalculatetheprobability,soanalystsdonothavetoexecutetheshowntransformationthemselves.Figure4.1.27showsthelogisticregression output for the response on a mailing example (as discussed in the CHAIDsection).Themodeloutputshowssomesignificantcoefficients(B)(sig.< .05):distancetothestore,typeofbank,andhouseholdsize.Itappearsthatcustomerswhoarefurtherawayfromthestorearelesslikelytorespondtothemailing.However,largehouseholdsaremorelikelytorespond.TheoutputalsoreportsthestandarderrorandtheWaldstatistic.ThesizeoftheWaldstatisticalsoshowswhichvariablehasthelargestimpactonanevent.Inthiscasedistancetothestorehasthelargestimpact,withaWaldstatisticof6.837.

Figure4.1.27OutputoflogisticregressionmailingexampleinSPSS

Computersciencemodels

In computer science also several models have been developed aiming to improvepredictions of the models discussed above. One problem with these models is that theresults might be affected by the specific composition of a sample. Therefore so-calledaggregation methods are used. The key idea behind these models is that improvedpredictionscouldbeobtainedbyaveraging the resultsofa largenumberofmodels.Theintuitionbehindaggregatingmultiplemodelresultsisthatthequalityofasinglepredictormight depend heavily on the specific sample (Breiman, 1996b) and is not knownbeforehand.Averagingpredictorsof varyingqualitywill result inmore stablepredictors(Breiman,1996a,Malthouse&Derenthal,2008).

An aggregation method that originated in the machine-learning field is bootstrap

aggregation,or“bagging”(Lemmens&Croux,2006).Inthebaggingprocedureamodelisestimatedonanumberofbootstrapsamplesoftheoriginalestimationsample,resultinginanumberofpredictionsforeverycustomer.Thefinalpredictionisobtainedbytakingtheaverageofallpredictions(Breiman,1996a).Thebaggingprocedureseemsespeciallyusefulfor improving the performance of decision trees. The performance of logistic regressionmodels is not substantially improved by using the bagging procedure, as the logisticregressionmodel results are less affected by sample composition (Risselada, Verhoef, &Bijmolt,2010).FormorespecificdetailsonbaggingwerefertoLemmensandCroux(2006).We also note that standard bagginghas not been implemented yet in standard softwarepackagessuchasSPSSandSAS.

Predictiveperformancemeasures

The literature proposes several metrics that can be used to assess the predictiveperformanceofpredictionmodels.Themostimportantare:

HitrateTop-decileliftGinicoefficient.

Hitrate

Thehitrateisdefinedasthepercentageofcustomersthatiscorrectlypredicted.Itcanbeformallyexpressedas:

The hit rate is a fairly easy metric to interpret, and is also a standard calculation. Animportantproblem,however, is that especially inmany customerpredictionmodels it isless informative due to the infrequent occurrence of events. Inmanymodels customersbecome mainly classified in the most occurring event. As a consequence it is almostimpossibletocomparedifferentcompetingmodelsandtoassesstheirpredictivequality.

Top-decilelift

“Lift” is the most common measure for model performance (Blattberg et al., 2008).Intuitively, lift measures whether the model is better able to distinguish between, forexample, responders and non-responders, or churners and non-churners. For thecalculationoftop-decileliftagains-tableneedstobedeveloped.Customersareassignedto

decilesbasedontheirpredictedevent(e.g.churn)probability.Foreachdeciletheaveragepredictedeventprobabilityiscalculated.Theliftissubsequentlycalculatedas:

InTable4.1.2weprovideanexampleof sucha calculation foramodelpredictingdirectmailresponse.Thetop-decileliftistheliftindecile1.Thetop-decileliftissubstantial,asthe response rate is 6%, while the average response rate is only 1.60%. Based on thecalculationsoftheliftsineachdecile,thecumulativeliftcanbecomputed.Thecumulativeliftofthekthdecileisdefinedbythepercentageofallrespondersaccountedforbythefirstkdeciles(Blattbergetal.,2008:319).InTable4.1.2thetoptwodecilesaccountfor59.4%ofallresponders.Thehigherthecumulativeliftforaspecificdecile,thebetterthemodel.

Ginicoefficient

The Gini coefficient is related to the lift model performance measures, in that it isessentiallytheareabetweenthemodel’scumulativeliftcurveandtheliftcurvethatwouldresultfromarandomprediction.Forthispurposeoneshouldcreateagainschart.Inthisgainschartthepercentageofcustomers,orderedbytheireventprobability,isplottedalongthex-axis,andthecumulativenumberofcustomersexperiencingtheeventisplottedalongthe y-axis (compare with cumulative lift). An example is shown in Figure 4.1.28. Thestraight line or the perfect equality line is the line representing a model with randompredictions(noinformation).Thecurveresultsfrompredictions.Alargerdistancebetweenthecurveandtheperfectequalitylinemeansastrongermodelperformance.

Table4.1.2Gainsandliftscale

Decile PredictedResponseRate Lift CumulativeLift(in%)

1 6.00 3.75 37.52 3.50 2.19 59.43 2.50 1.56 75.04 1.50 0.94 84.45 1.00 0.63 90.66 0.65 0.41 94.77 0.50 0.31 97.88 0.19 0.12 99.09 0.12 0.08 99.810 0.04 0.03 100.0Total 1.60 0.03

Source:AdaptedfromBlattbergetal.(2008)

Figure4.1.28Gainschart

TheGinicoefficientcanbecalculatedfromthegainschartasshowninFigure4.1.28.Wemainlyconsider thearea (A)betweentheperfectequality lineandthemodelcurve.Thecoefficient is calculatedbydividingAby theareaabove theperfect equality line (A+B).TheGinicoefficientcanhaveavaluebetween0and1.Avalueof0impliesthatthemodelpredicts just aswell as a random predictor. If themodel has a value of 1 it principallymeansthatthereisonlyonecustomerandthatcustomerisclassifiedasthecustomerwiththehighesteventprobability.Thisisaveryrarecase.Ingeneral,onewouldchosemodelsthathavecoefficientscloserto1.

PracticalIssues

Whenexecutingadatabaseanalysisonecanencountermanyissuesthatcanmakethelifeofadatabaseanalystprettytoughandchallenging.Belowwediscusssomeoftheseissues,which include problems with infrequent events, multi-faceted behavior, self-hiddenfindings,andsampleselection.

OccurrenceofInfrequentobservations:Indatabasemarketingitoftenhappensthatspecificeventsoccurinfrequently.Forexample,only2%ofcustomersmayrespondtoamailingoronly1%adoptanewservice.Topredictthiseventlargedatabasesarerequired.Aruleofthumbisthatananalystrequiresaround100customersforwhom the event occurs. This would imply a database of, at a minimum, 10,000customers for an event probability of 1%. This should not be too problematicnowadays.However, largedatabasesaremoredifficult tohandleandusuallyalsorequire more time for calculations to be carried out on them. One solution toovercomethesehurdlesistocreateabalancedsample,inwhichthecustomerswithaninfrequentlyoccurringeventareover-sampledandthecustomerswithnoevent

areunder-sampled. For the logitmodel this implies that only the constant in themodelisaffected,whiletheothercoefficientsofthepredictorvariablesshouldnotbe affected. The advantage of oversampling is that smaller databases can beanalyzed.Onemustbecarefultonotethatoversamplingleadstobiasedestimateswhenitisdoneinanon-randomfashionandthesampleconsistsofrelativelymanycustomers experiencing the event (e.g. churn) with a certain (influential)characteristic (e.g. many young customers). Donkers, Franses and Verhoef (2003)havereflectedonthisissueandhaveprovideddetailsonhowtocorrectforover-sampling.Multi-facetedcustomerbehavior:Inmanyinstancesthemodeledbehaviorisnotasingle event. For example, if a charity institution aims to predict response to amailing, they are also interested in the donation amount, as this also drives therevenueonthemailing.Inthesamevein,catalogersmightnotonlybeinterestedinthepredictionofapurchase,butalsointhesubsequentproductreturnprobability(Petersen&Kumar,2009).Inmanyinstancesbehaviorismulti-faceted.Fortunately,onecanapplymodelstopredictthesedifferentthoughrelatedbehaviorsjointly.Forexample,tomodelbothresponsetoacharitymailingandthedonationamount,onecan use a Tobit(2) model or one can model the purchase probability and thesubsequent order size (e.g.Konuş,Verhoef,&Neslin, 2014). In thismodel a logitmodelisusedtopredictresponseorpurchase,whilearegressionmodelisusedforpredicting thedonationamount (for responders)ororder size (forbuyers). In thesecondanalysisoneaccountsfortheselectioneffectofrespondingandthepotentialinterdependencebetweenthesetwoevents.Churn is not observed: In contractual relationships churn and many otherbehaviorsaretypicallyobservedinthedatabase,ascustomershavetorenewtheircontract. However, for many retailers one only observes buying behavior. Logitmodelsarethenlesssuitableforcustomeranalysis.Instead,durationmodelscanbeused tomodel the time to next purchase. Thesemodels can also be usefulwhenanalystsaim tomodel the timeuntil a customerchurnsoradoptsanewproduct(e.g. Prins & Verhoef, 2007; Polo, Sesé, & Verhoef, 2011). These models are alsoreferred to as hazardmodels, but go beyond the scope of this book.We refer toFransesandPaap(2001)foradetaileddiscussiononthesemodels.Self-hidden Easter eggs and more: The interactions between firms and theircustomers are often continuously recorded in a database.As a consequence, onecan observe the consequences of the interactions in real time. However, oneimportantissueisthatthisalsomayimpactthefindings.Let’sassumeforexamplethat customers with a high purchase volume usually receive a mailing. In asubsequent analysis, the researcher finds a link between purchase volume in thepast and purchase propensity. The question is whether this relationship occursbecauseofthehighvolumeorbecauseoftheselectivemailingstrategy.Inasensethe finding of this relationship could be labeled a self-hidden Easter egg in the

database. A related issue is that found relationships occur due to specificunderlyingmechanismswhichmightcausethefoundrelationshiptoappearstrongwhenitisactuallysmallerorevenabsent.Thisproblemoccurs,forexample,whenanalyzing the effect of loyalty programs (LPs) on customer loyalty. This effect isoverestimated,asonedoesnotaccountforthefactthat loyalcustomersaremoreinclinedtouseanLP.Ineconometrictermsonecallsthisanendogeneityproblem(seealsoBox4.1.2).

Conclusions

Thischapterhasdiscussedsevenanalyticalclassics.Theseclassicsvaryinapplicationarea,usedifferentmethodsandtechniquesthatcanbedividedintothedescriptivevs.predictivenatureandthestaticvs.dynamicfocus.Thesemethodsdo,however,haveincommonthattheyarefrequentlyusedwithinmarketingintelligencefunctions.Moreover,manyofthesemethodswill also play a role in big data environments.We believe it is therefore veryimportant to master these methods before immediately doing all kinds of big dataanalytics. Inthischapterwehaveaimedtoprovideacombinationofa theoreticalsounddiscussion of the differentmethods and important practical applications and insights ofthese methods. If readers are more interested in specific details, we refer to suggestedreadings cited in the text. Especially for analysts aiming to be an expert in specificmethods,understandingthedetailcanbeveryuseful.However,onewarningiswarrantedhere.Insomecasesanalystsmaybecometooexpert,withthepotentialdangerofignoringthepracticalimplicationsoftheir(advanced)modelingefforts.Insum,theanalystshouldbeuse-orientedwhenapplyingthedifferentmodels.

Notes

1WethankMarcelTemminghoffromGfKforsharingthesedata.

2WerefertoWedelandKamakura(2000)foranextensivediscussiononclusteranalysis.SeealsoHair,Black,Babin,

AndersonandTatham(2006)formoreonmultivariatedataanalysis.

3Thenumberofclusterscanalsobedeterminedwithsomestatisticalcriteria,suchasthefusioncoefficient.However,

thissubjectgoesbeyondthescopeofthisbook.Inanycase,thedendrogramistypicallyratherinsightful.

4Discriminantanalysis isamultivariate techniqueused toexplainandpredictgroupmembership.This technique is

availableinpackages,suchasIBM/SPSS.WerefertoHairetal.(2006)foramorein-depthdiscussion.

5We refer to Franses, vanDijk andOpschoor (2014) for an extensive discussion on time-seriesmodels. For readers

interested inmarketingtheworkofKoenPauwelswithmultipleapplicationsofseveralvariantsof theVARmodel

couldbeespeciallyrelevant.ForanoverviewoftimeseriesmodelsinmarketingwerefertoDeKimpeandHanssens

(2010).

6WethankFelixEggersforsharingthisexample.

7The levelassociation isusuallyassessedwiththeChi-squarestatistic,whichyoucanfind inaCHAID(Chi-square

automaticinteractiondetection)analysis.

8Thiscontestwasorganizedby theTeradataCenter forCustomerRelationshipManagementat theFuquaSchoolof

BusinessatDukeUniversityin2003.

References

Blattberg,R.C.,Kim,B.-D.,&Neslin,S.A. (2008).Databasemarketing–Analyzingandmanagingcustomers.NewYork:SpringerScience&BusinessMedia.

Breiman,L.(1996a).Baggingpredictors.MachineLearning,24(2),123–140.

Breiman, L. (1996b). The heuristics of instability in model selection. The Annals ofStatistics,24(6),2350–2383.

DavidShepardAssociates. (1999).Thenewdirectmarketing:How to implementaprofit-drivendatabasemarketingstrategy.Europe:Mcgraw-HillEducation.

Dekimpe,M.G.,&Hanssens,D.M. (1995).Thepersistenceofmarketingeffectsonsales.MarketingScience,14(1),1–21.

Dekimpe,M.G.,&Hanssens,D.M.(2010).Time-seriesmodelsinmarketing:somerecentdevelopments.MarketingJournalofResearchandManagement,1(1),93–98.

Donkers, B., Franses, P.H., &Verhoef, P. C. (2003). Using selective sampling for binarychoicemodels.JournalofMarketingResearch,40(4),492–497.

Donkers,B.,Verhoef,P.C.,&DeJong,M.G.(2007).PredictingCLV:Atestofcompetingmodelsintheinsuranceindustry.QuantitativeMarketingandEconomics,5(2),163–190.

Ebbes, P., Papies, D., & Van Heerde, H. J. (2011). The sense and non-sense of holdoutsamplevalidationinthepresenceofendogeneity.MarketingScience,30(6),1115–1122.

Eggers,F.,&Sattler,H. (2011).Preferencemeasurementwithconjointanalysis.Overviewof state-of-the-art approaches and recent developments. GfK Marketing IntelligenceReview,3(1),36–47.

Fader,P.S.,Hardie,B.G.,&Lee,K.L. (2005).RFMandCLV:Using iso-valuecurves forcustomerbaseanalysis.JournalofMarketingResearch,42(4),415–430.

Franses,P.H.,&Paap,R. (2001).Quantitativemodels inmarketing research.Cambridge,UK:CambridgeUniversityPress.

Franses, P.H., vanDijk,D.,&Opschoor,A. (2014).Time seriesmodels for business andeconomicforecasting.CambridgeUK:CambridgeUniversityPress.

Gijsenberg,M.J.,VanHeerde,H.J.,&Verhoef,P.C.(2015).Lossesloomlongerthangains:Modeling the impact of service crises on customer satisfaction over time. Journal ofMarketingResearch,52(5),642–656.

Hair, J. F., Black,W. C., Babin, B. J., Anderson, R. E., & Tatham, R. L. (4th ed.). (2006).Multivariatedataanalysis.NewJersey:PrenticeHall.

Hunneman,A.,Verhoef,P.C.,&Sloot,L.M.(2015).Theimpactofconsumerconfidenceon

storesatisfactionandshareofwalletformation.JournalofRetailing,91(3),516–532.

Konuş,U.,Verhoef,P.C.,&Neslin,S.A.(2008).Multiannelshoppersegmentsandtheirantecedents.JournalofRetailing,84(4),398–413.

Konuş,U.,Verhoef,P.C.,&Neslin,S.A.(2014).Theeffectofsearchchanneleliminationonpurchaseincidence,ordersizeandchannelchoice.InternationalJournalofResearchinMarketing,3(1),49–64.

Leeflang, P. S. H,Wieringa, J. E., Bijmolt, T. H. A., & Pauwels, K. H. (2015).Modelingmarkets.NewYork:Springer.

Leeflang, P. S. H., & Naert, P. A. (1978). Building implementable marketing models.Dordrecht:Kluwer.

Leenheer,J.,Bijmolt,T.H.A.,VanHeerde,H.J.,&Smidts,A.(2007).Doloyaltyprogramsreallyenhancebehavioral loyalty?Anempiricalanalysisaccounting forself-selectingmembers.InternationalJournalofResearchinMarketing,24(1),31–47.

Lemmens, A., & Croux, C. (2006). Bagging and boosting classification trees to predictchurn.JournalofMarketingResearch,43(2),276–286.

Lilien, G. L., & Rangaswamy, A. (2nd ed.). (2006). Marketing engineering: Computer-assistedmarketinganalysisandplanning.Victoria,BC,Canada:TraffordPublishing.

Lilien,G.L.,Kotler,P.,&Moorthy,S.(1992).Marketingmodels.PrenticeHallPTR.

Lindell,M.K.,&Whitney,D.J.(2001).Accountingforcommonmethodvarianceincross-sectionaldesigns.JournalofAppliedPsychology,86(1),114–121.

Malhotra, N. K. (2010),Marketing research: An applied orientation. New York: PrenticeHall.

Malthouse,E.C.,&Derenthal,K.M.(2008).Improvingpredictivescoringmodelsthroughmodelaggregation.JournalofInteractiveMarketing,22(3),51–68

Mark,T.,Lemon,K.N.,Vandenbosch,M.,Bulla, J.,&Maruotti,A. (2013).Capturing theevolution of customer–firm relationships: How customers become more (or less)valuableovertime.JournalofRetailing,89(3),231–245.

Neslin, S. A., Gupta S., Kamakura, W. A., Lu, J. X., & Mason, C. H. (2006). Defectiondetection: Measuring and understanding the predictive accuracy of customer churnmodels.JournalofMarketingResearch,43(2),204–211.

Netzer, O., Lattin, J. M., & Srinivasan, V. (2008). A hidden Markov model of customerrelationshipdynamics.MarketingScience,27(2),185–204.

Nijs,V.R.,Dekimpe,M.G.,Steenkamp,J.-B.E.M.,&Hanssens,D.M.(2001).Thecategory-demandeffectsofpricepromotions.MarketingScience,20(1),1–22.

Paas, L. J., Bijmolt, T. H. A., & Vermunt, J. K. (2007). Acquisition patterns of financial

products:Alongitudinalinvestigation.JournalofEconomicPsychology,28(2),229–241.

Pauwels, K. H. (2004). How dynamic consumer response, competitor response, companysupport and company inertia shape long-term marketing effectiveness. MarketingScience,23(4),596–610.

Petersen,J.A.,&Kumar,V.(2009).Areproductreturnsanecessaryevil?Antecedentsandconsequences.JournalofMarketing,73(3),35–51.

Podsakoff,P.M.,MacKenzie,S.B.,Lee, J.Y.,&Podsakoff,N.P. (2003).Methodbiases inbehavioral research: A critical review of the literature and recommended remedies.JournalofAppliedPsychology,88(5),879–903.

Polo, Y., Sesé, F. J., & Verhoef, P. C. (2011). The effect of pricing and advertising oncustomer retention in a liberalizing market. Journal of Interactive Marketing, 25(4),201–214.

Prins,R.,&Verhoef,P.C.(2007).Marketingcommunicationdriversofadoptiontimingofanewe-serviceamongexistingcustomers.JournalofMarketing,71(2),169–183.

Risselada,H.,Verhoef,P.C.,&Bijmolt,T.H.A.(2010).Stayingpowerofchurnpredictionmodels.JournalofInteractiveMarketing,24(3),198–208.

Roberts,J.H.,Kayandé,U.,&Stremersch,S.(2014).Fromacademicresearchtomarketingpractice: Exploring the marketing science value chain. International Journal ofResearchinMarketing,3(2),127–140.

Rossi,P.E. (2014).Eventherichcanmakethemselvespoor:Acriticalexaminationof IVmethodsinmarketingapplications.MarketingScience,34,655–672.

VanDoorn,J.,&Verhoef,P.C.(2008).Criticalincidentsandtheimpactofsatisfactiononcustomershare.JournalofMarketing,72(4),123–142.

Verhoef, P. C., Hoekstra, J. C., & Van der Scheer, H. R. (2009).Competing on analytics:StatusquovancustomerintelligenceinNederland.ReportofCustomerInsightsCenter(RUGCIC−2009–02),UniversityofGroningen.

Verhoef,P.C.,Spring,P.N.,Hoekstra,J.C.,&Leeflang,P.S.H.(2002).Thecommercialuseof segmentation and predictive modelling techniques for database marketing in theNetherlands.DecisionSupportSystem,34(4),471–481.

Verhoef, P. C., Van Doorn, J., & Dorotić, M. (2007). Customer value management: anoverviewandresearchagenda.MarketingJournalofResearchandManagement,3(2),105–120.

Wedel,M.,&Kamakura,W.A.(2000).Marketsegmentation:Conceptsandmethodologicalfoundations.Boston:KluwerAcademicPublishers.

Wittink, D. R., Addona, M. J., Hawkes, W. J., & Porter J. C. (2011). SCAN*PRO®: The

estimation,validation, anduseofpromotional effects scannerdata. InWieringa, J.E.,Verhoef,P.C.,&Hoekstra J.C. (Eds),LiberAmicoruminhonorofPeterS.H.Leeflang,UniversityofGroningen.

Wuyts,S.,Verhoef,P.C.,&Prins,R. (2009).Partnerselection inB2Binformationservicemarkets.InternationalJournalofResearchinMarketing,26(1),41–51.

4.2Bigdataanalytics

Introduction

Thedevelopmentsinbigdataarechanginganalytics.Therearethreemajordevelopmentsunderlying these changes. First, new data types and specifically non-structured data arebeinganalyzed.Traditionaldataanalyststypicallydonotalwayshavetheskillstoanalyzethese data, as totally new methods are required. Further, new data may require morecomputer science techniques.Second, in thisnewbigdataanddigital environment,newchallenges and questions arise.An important challenge, for example, is how to evaluateinvestments in new online advertising tools such as search engine advertising andaffiliates. Third, new analytical techniques are being developed that can account for thehugecontinuousdatainflow.Asaconsequence,atotalnewplayingfieldforanalystshasunfolded.This implies thattraditionalanalystshavetoadapttothesenewcircumstancesandhave tounderstandandbeable toapply thesenewtechniques.Fortunately, someofthesebigdataanalyticsstillusesomeoftheclassicsasdiscussedinChapter4.1.However,sometechniquesarerathernewandarenotincludedinthestatisticaltoolkitoftraditionalanalysts. In this chapterwe aim to discuss sevennewbig data analyses. These big dataanalysesareacombinationofnewtechniques,specificmarketingapplicationsandspecifictypesofdata.Notalloftheminvolveverysophisticatedmodels.InTable4.2.1weprovidean overview of seven new big data analytical areas, their importance for marketingdecisionsatthemarket,brand,orcustomerlevelandthestatisticalmethodstheyuse,andwewillgoontodiscusseachoftheseareasinmoredetail.

Bigdataarea1:Webanalytics

Web analytics can be viewed as a newbig data application area of standard descriptivetechniques, as it involves new data sources, which are typically more massive thanstandard data sources. Web analytics gained attention early in 2000, as early onlinemarketers tried to understand how consumers usewebsites and visit online stores. Thisbehaviorisdifferentfromhow,forexample,customersvisitsupermarkets(seeTable4.2.2).

Analysisofclickstreamdata

Toanalyzebehavioronwebsites, so-called “clickstreamdata” areused.Clickstreamdataare defined as the electronic record of Internet usage collected byweb servers or third-party servers. Two types of clickstream data can be distinguished (Bucklin & Sismeiro,2009):

Site-centric:detailed recordsofwhatvisitorsdowhennavigatingand interactingwith a specific site (offline: compare to loyalty card information of one specificstore)User-centric: detailed records of online behavior tracing across sites (offline:scannerdataofspecificconsumerproducts).

Basic analyses of clickstreamdata focus on thepurchase funnel on thewebsite.Websiteanalyticsaimtocollectmetricson thephaseof thepurchase funnel (seeFigure4.2.1). Inthisonlinefunneldifferentstepsarebeingmade,andonlinemarketsaimto improvetheusabilityofthewebsiteanduseothermarketinginstrumentstoincreasetheperformanceon each of these steps to finally end upwith a high conversion rate. In the end, theseanalysesarefairlydescriptiveandfocusoncounting.

Onecouldgobeyondcountingandconsiderwhatdrivesspecificbehavioronwebsites.For example, one could model the duration of a visit on the website and its effect onconversion (e.g. Johnson, Bellman, & Lohse, 2003; Bucklin & Sismeiro, 2003). Or theevolution of purchase behavior onwebsites can be considered (Shi& Zhang, 2014). Foronline retailersalsomarket-basketanalysis canbedone,which focusesonwhat isbeingpurchased during a visit and how the basket can be filled through personalized offers.Goingbeyondwebsiteanalytics,onlineandofflinemarketershavebecomeveryinterestedinthecustomerjourneyandhowthataffectspurchasebehavior—somethingwhichwewilldiscussinthenextsection.

Table4.2.1Sevenbigdataanalytics

Table4.2.2HowInternetchoicediffersfromsupermarketchoice

Internetchoice Supermarketchoice

•Intentunclearatoutset     Browse?Search?Buy?•Active/Interactive     Visitorparticipatesincreatingchoicecontext•Addressable     Choicecontextispersonalizable•Dynamic     Marketerscaninterveneatlowcost

•Storevisitrevealspurchaseintent•Passive•Fixed     Choicecontextcommon•Static

Source:Bucklinetal.(2002)

Largescaleexperimentation:A/Btesting

Figure4.2.1Onlinepurchasefunnel

Havinggainedanunderstandingofthepurchasefunnelandanyspecificissuesconnectedwithit,firmswillwanttogoonandsolvetheseissues.Thiscouldbedoneby,forexample,improvingsearchengineoptimization(SEO).Anotherwayistoimprovethewebsite.Firmsnowadaysuselarge-scaleexperimentationtotestchangesinwebsitedesign.Forexample,Booking.com changed specific wordings in offers and observed how this improved (orfailedtoimprove)conversionrates.Thislarge-scaleexperimentationisalsoknownasA/Btesting (see Figure 4.2.2). This A/B testing is nothing more than a randomized fieldexperimentamonga largenumberofvisitorsof thewebsite.Typically, anewwebsite istestedamongagroupofvisitorsand theresultsarecomparedwithacontrolgroupwhostill use the old version of the website. One can extend this testing by using morecomplicated experimental designs, such as factorial designs. If the experiment is fullyrandomized,ananalysisofvariance(ANOVA)canbeusedtotestforsignificanteffectsonoutcomes. To account for potential customer specific effects, such as age and loyalty,covariates can be included in an analysis of covariance (ANCOVA).Note that given thelarge scale of these experiments, not only the significance of these effects should beconsidered,butalsothesizeofthefounddifferences(seeChapter4).IngeneralA/Btestingisnotmuchdifferentfromthetestingofmarketingtacticsusedinpreviousyears,suchastestingdifferentversionsofdirectmailings(e.g.DavidShepardAssociates,1999).ProbablytheonlydifferenceisthatA/Btestingusuallyinvolvesverylargesamples.

Figure4.2.2A/Btesting

Bigdataarea2:Customerjourneyanalysis

Customer journey analysis considers how customers interact with multiple touchpointsmoving from consideration, search, and purchase to consumption and after-sales. Theresearch is interested in describing this journey and understanding the choices fortouchpointsinmultiplepurchasephases.

The extended customer journey analysis typically uses qualitative interviews, surveydataorobservations. Inservicemarketingblueprintinghasbeendevelopedtoassess theimportanceofseveral touchpoints intheservicedelivery(e.g.Bitner,Ostrom,&Morgan,2008).Ithasalsobecomeanimportanttopicinmulti-channelresearch(Verhoef,Kannan,&Inman, 2015). Customer journey analysis is frequently executed using qualitativetechniquesinwhichcustomersareaskedtodescribetheirchannelsandtouchpointsusedand which channel they prefer to use. Based on these qualitative insights firms maydevelopstrategiesonhowcustomerscanoptimallybesteeredthroughtheirchannels.

Amorequantitativeanalysisaims todescribe theusedchannels indifferentphasesofthepurchaseprocess.Typically,onemakesadistinctionbetween(Neslinetal.,2006):

Channels/touchpointsusedforsearchChannels/touchpointsusedforpurchaseChannels/touchpointsusedforafter-sales.

Thechannelscanbebothofflineandonline.Channelbehavioronlinecanbetrackedmoreeasily (using cookies) than tracking channel usage offline.Onlywhen customers can beidentifiedoffline(e.g.throughaloyaltycard)mightonebeabletoobservetheirbehaviorand store that in a database.Note that even in this case one only observes the channelusageforcustomersthatvisitedtheofflinechannelofthefirmandpurchased.Thismeansthat, for example, avisit toa store to search (butnotpurchase) isnotobserved,whileapurchaselinkedto,forexample,aloyaltycardisobserved.Hencetogetafullpictureofthechannel of usage of customers firms frequently have to rely on survey datawhich askswhichchannelshavebeenusedforsearch,purchase,andafter-sales(e.g.Verhoef,Neslin,&Vroomen,2007;Gensler,Verhoef,&Böhm,2012).Surveydatamainlyrecallchannelusageand/orchannelpreferences.

Specific household panels of firms likeGfKmay also observe this behavior, but thenfrequently only purchase data are recorded (e.g.Melis, Campo, Breugelmans, & Lamey,2015).Arelativelynewmethodisreal-timetrackingusingmobiletechnologies.Customerscan use a mobile app to record their interactions with multiple brand touchpoints(MacDonald,Wilson,&Konuş, 2012). It is likely that this techniquewilldevelop furtherand that based on location data and amobile app, visits to stores etc. will be recordedautomatically. Still, this will only occur for a sample of customers—those who allowresearch firms toobserve theirbehavior. InFigure4.2.3we report a survey studyon the

effectsofdifferent touchpoints thatcustomershavebeenexposed toand theireffectsonadvertising recall and brand consideration. Digital and television jointly have a strongeffect on brand consideration and thus jointly contribute to the effectiveness of thecampaign.

The analysis of these channel usage data typically results in a description of channelusagepatterns.Reportsmightalsodiscusstheuseofmobilephoneswhenshoppingofflineetc.Soinessencemostoftheseanalysesarefairlydescriptiveandcannotbelabeledasbigdata analytics. In Figure 4.2.4 we provide some of these statistics as an example andspecificallyontheshowroomingvs.webroomingbehavior.Showroominginvolvesvisitingthe store for search and subsequently switching to the online channel for purchase,whereaswithwebroomingtheonlinechannelisusedforsearchandthestoreisusedforpurchase(e.g.Rapp,Baker,Bachrach,Ogilvie,&Beitelspache,2015;Verhoefetal.,2015).

Customer segments on channel usage can also be distinguished. For example, Konuş,VerhoefandNeslin(2008)distinguishbetweenenthusiasticmulti-channelshoppers,store-shoppers, and unenthusiastic multi-channel shoppers, using latent class analysis as asegmentation technique (see Figure 4.2.5). These segments, however, differ between thestudiedproductcategories.

Figure4.2.3Effectofdifferenttouchpointsonadvertisingrecallandbrandconsideration

Figure4.2.4Useofdifferentchannelforsearchandpurchase:Webroomingvs.showrooming

Source:AdaptedfromEdwards(2014)

Figure4.2.5Latentclasssegmentationbasedoncustomerchannelusage

Source:AdaptedfromKonuşetal.(2008)

Asanextsteponemightaimtoexplainand/orpredictchannelchoicesinthedifferentphases. To explain channel usage in different phases, one could include customercharacteristics(e.g.age,income),customerbuyingbehavior(e.g.customervalue,purchasefrequency), attitudes towards channels andprior channelusage (e.g.Verhoef et al., 2007;Konuş, Neslin, & Verhoef, 2014). Typically choice models are used. This becomes morecomplicatedwhenmultiple channels canbeused simultaneously,which typically occursforthesearchphase(i.e.onlineandstoreforsearch).Multivariateprobitmodelscanthenbeapplied,as thesemodelsallow formanychoices tobemade. In thepurchasephaseasingle channel should be chosen and then logit or probit models can be used. Channelchoicesalsodependoneachother,through,forexample,channellock-in.Channelchoicesinthesearchphasemayaffectchannelchoicesinthepurchasephase,andviceversa(e.g.Gensler et al., 2012). In principle the usage of channels can be considered as a kind ofswitching process between channels during different phases. In that sense, Markov

switchingmodelscouldbeuseful.However,sofarwearenotawareofstudiesusingthismethod.

Asanextstep,channelchoicescanalsobelinkedtopurchasebehavior.Forexample,onequestion might be whether a channel migration (e.g. moving from catalog to online)increasesthepurchaseprobability(e.g.Ansari,Mela,&Neslin,2008),orwhetherchanneleliminationreducestherevenuesfromcustomers(seeFigure4.2.6,andformoreextensivemodelresultsseeKonuşetal.,2014).

Figure4.2.6Revenues,costs,andprofitpergroupwithandwithoutsearchchannelcatalog

Source:AdaptedfromKonuşetal.(2014)

Bigdataarea3:Attributionmodeling

When purchasing online, customers go throughmany stages according to the purchasefunnelapproach.Customersareinfluencedbyseveraltouchpoints.Thesetouchpointscanbe firm initiated—for example via emails—or customer initiated—for example, via searchengines (De Haan, Wiesel, & Pauwels, 2013) and may occur at different phases of thepurchasefunnel.Customersmaythusvaryintheirreadinesstopurchase(seeFigure4.2.7).Firmswilltypicallyonlyhavelimitedinformationonthetotalusageoftouchpointsinthis“path to purchase” and will typically only observe the ones that are used to visit theretailer’swebsite.Consequently,thequestioniswhetherthesalecanbeattributedtothislast-used touchpoint or to other touchpoints, which are generally not observed. It isessentialtoknowthis, forallocatingresourcesamongtouchpoints. It isalsoimportanttomakedealswithpartners,suchasGoogleorcomparisonwebsites.Toassesstheeffectsofcustomer touchpoints on conversion and sales, attribution modeling can be used.Sometimesthisisalsoreferredtoasthedevelopmentof“pathtopurchase”models.

Methods

Themostcommonmethodforattributingsalesis“last-click.”Inthisrathernaivewaythesaleisattributedtothelastusedtouchpoint.Thisisarathersimpleapproachanddoesnotrequire any modeling. To assess the performance of each touchpoint one can calculatemetricssuchasthenumberofwebsitesvisited

Figure4.2.7Purchasefunnel:Pathtopurchaseonmobilehandset

throughthistouchpoint,andtheconversionratepertouchpoint.Itisquestionablewhetherthisisarightapproach,asthetouchpointsusedbeforethelastonecontributedaswell.Solast clickwill typically overestimate the contribution of the touchpoint. There are othermethodsavailable. “First-click”attribution isused: thepurchase isattributed to the first-

used touchpoint. Average attribution methods assume that every used touchpointcontributes equally to the sale: this can be adjusted by using time-decay attribution byassuming that the latest used touchpoints contribute more strongly to the sale. Someadjustments can also be made by using other weighting scores. This is referred to as“customerattribution”(VanderHeijden,DeHaan,&Hoving-Wesselius,2014).

Amorecomplicatedmethodofattributingsalesinvolvescreatingadatasetinwhichthepurchase (or not) is recorded, along with the used touchpoints, and if possible othercustomercharacteristics.Alogisticregressionmodelisthenusedtoestimatetheeffectofeach touchpoint onpurchase.However, one concernhere is that in some channels somecustomershaveahigherinherentreadinesstopurchase.Thiswouldimplyendogeneityofthe touchpoints, and one should correct for this (see Chapter 4.1). This is notstraightforward, asmore information is required and this information is frequently notavailable. In a case studyon attributionmodelingpresented inChapter6we showhowtouchpoint usage in the last purchase occasion can function as away to correct for theendogeneity in touchpoint usage. And indeed the contributions to purchase of specifictouchpointssuchassearchenginesarecorrecteddownwards.Interestingly,thisstudyalsoshows that the results of correcting are rather similar to amodel that accounts for theeffects of touchpoints and some customer characteristics, such as customer loyalty. InFigure4.2.8weprovidetheestimatedeffectsofa“correct”attributionmodelandamodelinwhichthe lastclickassumptionisbeingused.Asonecanobserve, therearerelativelylargedifferencesbetweenthefoundeffects.

Figure4.2.8Comparisonofeffectsestimatedbyattributionmodelandlastclickmethod

These attributionmodelingmethods involve an individual level approach to assessingthe contribution of specific touchpoints. This implies that individual-level data oftouchpointsare required.Asnotedearlier, thesedataareavailable foronline retailersofthe last-used touchpoint.More elaborate searches could be available if cookies are used.Butofflinetouchpointsarenotavailable.Measuringtheeffectsofofflinetouchpointsattheindividuallevelrequiresthatcustomersareconstantlymonitoredinwhattheyseeanddo.Panelsarenowbeingdevelopedtocollatethisinformationforasampleofcustomers.Wehave already discussed specific methods using mobile panels (MacDonald et al., 2012).

Anothersolutionistouseaggregateddailydata,inwhichforanonlineretailerthedailyspendinginoffline(e.g.advertising)andonline(e.g.searchengine)mediaismonitoredandlinkedtothedailyonlinewebsitemetrics(i.e.websitetraffic,sales).DeHaanetal.(2013)describehowtheyusethesedatatoassesstheeffectsofonlineandofflinemediaspendingonthesemetrics.TheyuseaspecificversionoftheVARmodel(seeChapter4.1).Althoughwith this approach both offline and online media spending can be considered, onlinemarketersconsiderthisaggregateapproachtoberatherdifficultastheyaresomuchusedtoanalyzingindividual-leveldatawhendoingwebanalytics.

Bigdataarea4:Dynamictargeting1

Targetingcustomerswith the rightoffershasbeenaround fora long time. Inparticular,direct marketing models have been developed to select those customers who are mostresponsivetodirectmailings(e.g.Bult&Wansbeek,1995).ThearrivalandgrowthoftheInternetanddigitaldeviceswithappshavemoved targeting tohigher levels.Using real-time behavioral data firms aim to provide personalized offers to customers visiting thewebsiteorloggingontoanapp.Giventherelativelylowcostsofapproachingcustomersonline,personalizationstrategieshavebecomeeconomicallymoreattractive(e.g.Zhang&Wedel,2009).Successfulpersonalizationcanbeawaytogainacompetitiveadvantage,asitcouldresultinmoresatisfiedcustomersandmoreeffectivemarketing.Asaconsequence,closed loop marketing (CLM) has become popular. CLM consists of a cycle in whichcustomer information is continuously collected andupdated, and advanced analytics areusedtoforecastcustomerbehaviorandusedtoredesignandpersonalizeproducts,services,andmarketingeffort,inshortcycles(Chung&Wedel,2014),asshowninFigure4.2.9.

We distinguish between twomajor types of dynamic targeting approaches that differwithrespecttothemethodstheyuse(Chung&Wedel,2014):

RecommendationsystemsPersonalizationsystems.

Figure4.2.9Closed-loopmarketingprocess

Source:AdaptedfromChung&Wedel(2014)

Recommendationsystems

Recommendation systems have been around since the start of Internet retailing and areusedextensivelybyfirmssuchasAmazonandNetflix.Thekeyideaofrecommendationsystemsisthatbasedonacustomer’scharacteristicsandcharacteristicsofothercustomers,

specific recommendationscanbegiven tocustomers (e.g. “thisbookmightbesomethingforyou”).Threetypesofrecommendationsystemscanbedistinguished(Chung&Wedel,2014):

ContentfilteringsystemsCollaborativefilteringsystemsHybridformsofcontentandcollaborativefilteringsystems.

Content filtering systems involvedigital agents that produce recommendations basedonthe target customer’s past preferences for products/services and the similarities betweenthoseproducts/services.Henceproductsorservicesareofferedthatarerathersimilartotheonespurchasedbefore.Forexample,ifonefrequentlybuysfantasybooks,itislikelythatthenextrecommendationwillbeforafantasybookaswell.Collaborativefilteringaimstomake a recommendation using the preferences of other, similar customers. In practice,frequently so-called memory-based systems are being used here. These systems usemeasures of similarity between customers’ preferences or behaviors. These systems aresimpletoimplement,easilyscalable,androbust(Chung&Wedel,2014).However,whenitinvolvesbillionsofrecommendationsformillionsofproducts,moreadvancedtechnology,such asmap reducing, could be needed. InChapter6we describe the case of an onlineretaileractuallyusing this technology toachieve theirbusinessandmarketingobjectiveswith their recommendations. Another technique often adopted is the nearest neighbouralgorithm. Because of reported problems with the more simple memory-based systems,model-basedsystemshavebeendevelopedinthemarketingliterature(e.g.Bodapati,2008).Model-basedsystemsusing, forexample, stochastic,Bayesian,and/or latent-class typeofmodelstendtooutperformthememory-basedsystems,butarecomputationallyintensive(Chung&Wedel,2014).

Ansari,EssegaierandKohli(2000)suggestthatrecommendationsystemsshouldnotonlybe based on customers’ revealed preferences and the revealed preferences of othercustomers, but should also involve preferences for product attributes, expert judgments,and specific customer characteristics. Recommendation systems are therefore sometimesupdated with preferences for attributes. Firms can use conjoint analysis to deriveindividual estimates of attribute utilities (see Chapter 4.1). The problem with attributeutilitiesisthattheyarelikelytobecollectedonlyforalimitednumberofcustomers.Fornewcustomerstheseattributeutilitiesarenotavailableandthustheyshouldbecombinedwith behavioral patterns of customers. For example, a combined algorithm using bothconjointutilitiesandbehavioralpatternscouldbedeveloped,therebyalsoconsideringthebehaviorofothersimilarcustomers,forawebsitesellinghotelbreaks.InFigure4.2.10weprovide an overview of this algorithm. Importantly, this algorithm also accounts for thereviews written about the hotel. This is actually an ongoing trend in recommendationagents. Many recommendation agents nowadays use reviews written by customers, asmanycustomerswritereviews.ThepopularityofsocialnetworkssuchasFacebook,where

customerscan like productsandbrands, stimulates theuseof reviews in theseagentsaswell.

Personalization

Thekeydifferencebetweenrecommendationsystemsandpersonalizationsystems is thatrecommendation systems recommend existing products or services, whereaspersonalization systems adapt the offering to customers’ needs. Note that it not onlyinvolves product or service offerings—how the website is displayed (looks and feels) tocustomers can also be personalized. In that sense, personalization systems are morecustomer-driventhanrecommendationsystems.Customizationisatermthatisalsousedfrequently in the context of personalization.Customization involves the firm facilitatingthe customer by tailoring products and services to his or her preferences (Arora et al.,2008).AfamousexampleisDell.com,wherecustomerscancombinespecificattributestocustomizetheirpreferredcomputer.CarmanufacturerssuchasBMWalsoallowcustomerstocustomize.Thefirmsthusprovideaninterfacetocustomerstocustomizeaproduct:thecustomer ismakingdecisions.Withpersonalizationthefirmtailors themarketingmixtothecustomer,basedonavailablecustomer

Figure4.2.10Schematicoverviewofrecommendationagentinhotelindustry

information (Arora et al., 2008). Examples of personalization can be found in manyindustries and for some—especially for information-based services—personalization hasbecomekey.Forexample,Last.fmmightplaymusicbasedonacustomer’spriorselectionofsongsandlikes/dislikesofsongs.Pandorasuggestssongsonthesepriorselectionsaswellasonsimilaritiesbetweensongattributes(Chung&Wedel,2014).

Forpersonalizationsystemsitisveryimportanttolearnconsumerpreferences.Thiscanbedoneusingashortsurveyonconsumerpreferences.Basedontheresultsofthisexercise,

apersonalizedoffercanbeprovided.Andbasedontheoutcomeofthisoffer(i.e.acceptorreject offer) a consumer’s preferences canbeupdated (seeCLMapproach). Especially inmoremobileenvironments,adaptivepersonalizationsystems(APS)havebeendeveloped.APS takes full advantage of unobtrusively obtained customer information to providepersonalizedservicesinrealtime(Rust&Chung,2006;Chung,Rust,&Wedel,2009).Thesesystemsrequireverylimitedornoinputfromthecustomerandrelyheavilyonpurchasedata. Importantly,APScanlearnfromsmallpiecesof informationovertimeandpickupchangesinconsumerpreferences.

InthemarketingliteraturetherearemanyapplicationsoftheseAPS,whichvaryintheiruseofstatisticalandeconometricmethods.Oneimportantissueinpersonalizationisthatoneaimstoaccountforcustomerheterogeneityintheusedmodel.Severalpersonalizationsystemsusealatent-classtypeofmodels,forexampleafinitemixturemodel(e.g.Zhang&Wedel,2009),inwhichparametersforspecificcustomersegmentsareestimated(e.g.Wedel&Kamakura,2000; seealsoChapter4.1).Forotherpurposesspecificalgorithmsareusedthatfrequentlyhavetheirbackgroundinartificialintelligence.Forexample,Chung,Rust,& Wedel (2009) use a dynamic Markov Chain Monte Carlo method as a filteringmechanismandalsouseacombinationofmodel-basedandcollaborativepersonalizationapproaches.Other algorithms are used in other studies. The details of thesemethods gobeyondthescopeofthisbookandwerefertosomeofthecitedarticlespublishedintopmarketing journals. One method to develop more individual estimates that has gainedattention is hierarchical Bayesian model estimation (e.g. Rust & Verhoef, 2005). WithBayesianmodel estimation individual regression parameters can be estimated. The verybasicandintuitiveprinciplesofmodelsaccountingforheterogeneityarediscussedinBox4.2.1. It isworthnoting that thismodelingapproach isnotonlyused forpersonalizationpurposes.

Box4.2.1Theessenceofmodelingcustomerheterogeneity:MovingtohierarchicalBayesianmodelsTraditionalregressionandchoicemodels inmarketingassumesimilareffectsforallincluded variables in the models. Assume for simplicity that we aim to explain achoice (i.e. response to an offer, new product etc.). This choice for individual i isrepresentedinalogitmodelusinganobservedlatentvariableYi*andweexplainthischoicebyonevariableprice.Inprinciplethesimplelogitmodelisthenformulatedasfollows:

As noted, both the constant and coefficient of price are similar for all consideredindividuals. This assumption is, however, not so plausible. Customerswill responddifferently to marketing mix instruments, including price. Moreover, the inherent

probabilityofchoosingabrandorproductmaydifferbetweencustomers.Inprinciplesome heterogeneity could be captured by including additional customercharacteristics,suchasage,income,usage,etc.Heterogeneitycanalsobeaccountedforbyusinginteractions.Forexample,aninteractionbetweenpriceandincomecouldbeincluded,ashigh-incomecustomerscouldbelesspricesensitive.Anextstepistoallowtheparameterstovarybetweensegments(s=1..S).Thisisactuallywhatfinitemixturemodelsdo.Thebasicmodelisthenformulatedasfollows:

Theparametersnowdifferbetweensegments.Thenumberofsegmentsisdeterminedbyconsideringfit-criteriaandtypicallydiffersbetweenapplications.IntheirstudyonpersonalizedpromotionsZhang&Wedel(2009)findtwosegments.

Hierarchical Bayesian models allow research to estimate individual parameters.Theseparametershaveadistribution,andcansubsequentlybeexplainedbyspecificothervariablesZ(i.e.,age,income,usage,relationshipduration).Formally,thiscanbewrittendowninahierarchicalwayasfollows:

To estimate this in a Bayesian way, simulations (e.g. Markov Chain Monte Carlo(MCMC)) should be done and specific samplers (i.e. Gibbs) should be used. Thisresults inmorecomplicatedequations thanwrittendownhere.Theresultsof theseanalysescanbeusedinpersonalizationandcanbeusedtodevelopoptimalone-to-onepersonalizedmarketingstrategies(e.g.Rust&Verhoef,2005).ForveryinterestedreaderswerefertoworkbyOhioStateprofessorGregAllenbyandUCLAprofessorPeter Rossi, who are the leading experts in Bayesian models in marketing(www.econ.umn.edu/~bajari/iosp07/rossi1.pdf;seealsoAllenby,Rossi,andMcCulloch(2005)).

Bigdataarea5:Integratedbigdatamodels

Oneofthekey-characteristicsofbigdataisthatitinvolvesmultipledatasourcesthatcanbecombinedtopredictmarketphenomena.Awell-knownexampleisthatthenumberofindividualswith flu can be predicted by looking at the number of specific searchers onGoogle(seeFigure4.2.11).Withthisthefluactivityaroundtheglobecanbeestimated.2

We derived this graph in Spring 2015.Google has since decided to no longer publishcurrent estimates of flu and dengue based on search patterns. Some bloggers seem toattributethisdecisiontocriticismoffluestimates.3

Overallinabigdataerawemovefromsingledatasourcestomultipledatasources,andthis may also involve new data sources. This may enrichmodels and in turn induce astrongerpredictivepower.However,oneshouldalsobecareful!Specifically,theGoogleflupredictions sometimes overestimated flu activity. Lazer, Kennedy, King, & Vespignani(2014)warnabouttheuseofthesenewdatasources.4Oneoftheirconcernsisoverwhattheycall“bigdatahubris.”Theywarnanalystsnottojustreplaceexistingdataandmodelswithnewsocialandonlinedata.Insteadofsubstitutingforolddataandmethods,newdataand models should complement them.Moreover, they warn that big data sources havemeasurementproblemsandshouldbeconsideredinasimilarwaytoolddata.Whatisthevalidity of the newdata?How is itmeasured andwhat is themeasurement error? (SeeChapters 3 and 3.1). And so on. For example, with the flu search data, one might beconcerned about endogeneity and causality. The fact that people are searching probablygoeshand-in-handwiththemhavingtheflu.

Figure4.2.11FluactivityUSApredictedbyGoogle

Source:Adaptedfromwww.google.org/flutrends/about/how.html

Havingmade these very important caveats on newdata sources, it is still undeniablethatbigdataallowsmoredatasourcestobeintegratedtodeveloprichermodelsthatarelikely to lead to a better understanding of marketing phenomena and predictions. Thisrequiresadeeperunderstandingof theanalyticalchallengesencounteredwith integratedmodels.However,itmayalsoprovideopportunitiesfortheapplicationofnewtechniques.

If one starts combiningdata sources, one encountersmanyproblems, such asmissingvalues.InChapters3.1and3.2wediscussedmanyoftheseissues,includingmissingvalues,andsolutionssuchasdatafusion.Fornow,weassumethatwehavesolvedtheseissuesandsufficient data are available. Before analyzing the data, it is important to assess theaggregationlevelofmultipledatasources.Istherethesameaggregationlevelintermsofanalyzed entity (e.g. market, brand, customer) and time (e.g. daily, weekly, monthly,yearly)?Ifaggregationlevelsaresimilar,whichcouldbethecasewhenpredictingweeklysalesofaspirinusingpastsalesofaspirin,trends,andweeklyGoogletrendsonflusearchkeywords,thestandardmodels(seeChapter4.1)canbeusedwithoutanycomplications.

Combiningdifferentaggregationlevels

Challengesmayoccurwhenfirmsstartcollectingdataofmultipleentities,forexample:

One aims to explain individual monthly customer retention for mobile phoneoperatorusingindividualvariables(e.g.age,usage,relationshipduration),butalsomonthly brand advertising data and weekly Google searches on mobile phoneplans.Oneaimstoexplainindividualmonthlycustomersatisfactionovertimeforapublictransport organization, using data on individuals, public transport usage, weeklypunctuality of used train-connections, and available services on visited trainstations.One aims to explain weekly brand sales using data on (competitive) brandadvertising, pricing, and promotions, individual daily brand likes and individualmonthlymeasuredbrandattitudes.

Somegeneralsolutionsareasfollows:

Aimtoaggregatethedatatoone(single)aggregationlevelandanalyzethedataatthataggregationlevel.Asimpleruleisthattheaggregationlevelofthedependentvariableisdecisiveinsuchawaythatonecannothaveexplanatoryvariableswithloweraggregationlevelsthantheaggregationlevelofthedependentvariable(e.g.abrand leveldependentvariablecannotbeexplainedby individual levelvariables).In the brand sales example, one would aggregate the data at the brand level.Individualbrand likes andattitudeswould thenbe aggregated to averageweeklybrandlikesandmonthlybrandattitudes.The time aggregation problem can also be solved by making assumptions. Forexample, for monthly brand attitudes, one might assume that the attitudes aresimilar for every underlying week in that month. Or one might assume sometransitionsbetweenthedifferentmonths.Thebrandattitudesinthefirstweekofamonth are then partially based on the attitudes of past month and the current

month.

Multi-levelmodels

Onepotentialmodelsolutiontoconsideristheuseofamulti-levelmodel.Inamulti-levelmodel one acknowledges that there are specific aggregation levels in the data analyzed.The classic example is the analysis of a children’s school performance.Children (lowestaggregationlevel)arepartofaclassorschoolandarelivinginaspecificregion(highestaggregationlevel).Inamarketingcontextindividualcustomersmightbuydifferentbrandsin many stores. In a multi-level model at different aggregation levels equations areformulated.Hereweaimtodiscussthebasicsofmulti-levelmodels.Weassumewehavetwolevelsinthedata:brands(j)andcustomers(i).Inourexampleweaimtoexplaintheloyaltyofcustomers tomultiplebrandsandtheroleofbrandattitudes (BrAtt)andsomecustomercharacteristics(X).Inamulti-levelmodelonefirststartswithamodelatthefirstlevel(customerlevel).

Atasecondlevelonemightassumethattheconstant(basebrandloyalty)andtheeffectofbrandattitudesisaffectedbysomebrandlevelvariables,suchasaveragebrandadvertising(Adv).Thefollowingequationsarethenformulatedtoaccountforthis:

Notethatwedonotassumethattheeffectsofsomecustomercharacteristicsareinfluencedbyadvertising,asthisparameterβ2issimilarforallbrands.Toapplyamulti-levelmodelitiscrucialtohavemultipleentities(i.e.brands)perlevel.Amulti-levelmodelisnotusefulifonlyonebrandisstudied—inthatcasethereissimplynovariationinadvertising.Formoretechnicaldetailsonmulti-levelmodelswerefertoSnijdersandBosker(2012).

We have also applied multi-level analysis in customer management applications.Specifically,whenconsideringtheeffectsofmultiplecustomerfeedbackmetrics(CFM)onretention (see Chapter 4.1), we used multi-level models, as we had observations ofcustomers ofmultiple brands inmultiple industries (DeHaan,Verhoef,&Wiesel, 2015).Hence,weusedamulti-levelmodelwiththree levels.Thisalsoallowedusto investigatetheeffectsofCFMsatthreelevels.Atthecustomerlevel(CFMcust.)westudiedwhetherahigherCFM score resulted in higher customer loyalty.At the firm level (CFM firm)weinvestigatedwhetherahigherCFMthanotherbrandsresultedinahigheraverageloyaltytoafirm.AttheindustrylevelweconsideredwhetherindustrieswithhigherCFMscoreshave a higher average loyalty. We estimated models per CFM. We also allowed forcorrelations between the different equations in themodel. The results for the customer

satisfactionmetricand theNPSmetricaredisplayed inFigure4.2.12 (seeDeHaanetal.(2015)formoredetails).Themajorsignificanteffectsaretypicallyfoundat thecustomerandfirm-level.

Adaptiveforecastingtechniques

In a broader sense, big data developments have also stirred up interest in adaptiveforecastingtechniques,astheyhaveenrichedtheunderstandingoftheideathatestimatedeffectsmightbeaffectedbyevents.Moreover,inadynamicworlditisunlikelytoberightto assume that parameters of models are stable. Especially for forecasting purposes,accountingforthesedynamicsiscrucial.Adaptiveforecastingintime-seriesanalysishadalready been developed in the 1960s to account for regular events, such as seasons (e.g.Mincer, 1969). Thereby, they assume that coefficients and error structures may varybetween these events. For example, a very simple assumption might be that duringthanksgivingorChristmas sales are expected tobehigher, andas a consequence in thatperiodtherelevantconstantinamodelshouldbeadjusted.

Figure4.2.12Estimationresultsofmulti-levelmodeltoassessperformanceofCFMs

Source:AdaptedfromDeHaanetal.(2015)**p<0.01,*p>0.05(one-tailedfortheCFMs,twotailedfortherest*Significantly(p0<.05)smallerthanthefirm-leveleffect

Another way to account for changing events is using so-called “regime switching”models (Hamilton, 2008). Under specific regimes (e.g. expansion of economy) specificparameters are estimated, while under other conditions (e.g. economic recession) otherparametersareestimated. It isofcoursecrucial tounderstandwhentheseeventchangesoccur. For forecasting purposes, onemight develop predictions under different scenarios(regimes),whichallows,forexample,thepredictionofhowsaleswoulddevelopunderbadandgoodeconomictimes.

Bayesianmodelscanalsobeusedtoaccount forchanges in theenvironment (seeBox

4.2.1). One way would be to allow parameters in the model to change over time. Forexample,theeffectsofmarketingspendingmaydependonthecompetitioninthemarketorthegrowthrateinthemarket.Ifthedatasetissufficientlylargeandsufficientvariationoccurs, the effects ofmarketing spendingmaydependon thesevariables and thus time-varyingmarketingeffectscanbeestimated.AnexamplecanbefoundinAtaman,Mela,&Van Heerde (2008) where time-varying effects of different marketing mix variables areestimated for newly introduced products for which it is reasonable to assume varyingparameter estimates (see Figure 4.2.13). They come up with the effects of differentinstruments for achieving growth and building market potential. As can be observed,achievingdistributionisofutmostimportance.

Bigdataarea6:Sociallistening

Figure4.2.13Effectsofmarketingmixvariablesonbrandperformanceusingtime-varyingparametermodels

Source:AdaptedfromAtamanetal.(2008)

Sociallisteninghasbecomeoneofthemostpopularnewanalysistools,asinthisbigdataeraunstructuredandspecificallytextdatahasbecomeomnipresentinonlinecommunities,social media like Facebook and Twitter, as well as in reviews on websites such asTripAdvisor, and retailers such asAmazon. These data are accessible tomultiple partiesand research agencies have specialized in providing firms with information on howcustomersviewtheirfirmsandbrands.Therearealsosomestandardtoolsavailable,suchas Radian6 for socialmediamonitoring. Text analytics is used for the analysis of socialdata.Textanalyticshasbeenaroundforyears,butithasnowbeenautomatedandfirmslike IBMandSASoffer tools toanalyze thesedata.A famousapplication isDr.Watson,developedbyIBM.Beyondtextanalytics,picturesandvideoscanalsobeanalyzed.Wewillmainlyfocusontextanalyticsinourdiscussion,asthisismostwidelyadoptednowandisconsidered an important big data analytics technique (Chen, Chiang, & Storey, 2012).According to Chen et al. (2012) text analytics have their academic roots in informationretrievalandcomputational linguistics. In informationretrieval,document representationand query processing are the foundations for developing a vector-spacemodel, Booleanretrievalmodel,andprobabilisticretrievalmodel,whichinturnbecamethebasis for themodern digital libraries, search engines, and enterprise search systems (Salton, 1989). Incomputational linguistics, statistical natural language processing (NLP) techniques forlexical acquisition, word sense disambiguation, part-of-speech-tagging (POST), andprobabilisticcontext-freegrammarsareofhugeimportanceforrepresentingtext(Manning&Schütze,1999).

Marketingresearchagencieshavealsospecializedintextanalytics.Itismainlyusedtounderstand opinions of customers, something which is often called “opinion mining.”Opinion mining refers to the computational techniques for extracting, classifying,understanding,andassessingtheopinionsexpressedinvariousonlinenewssources,social

mediacomments,andotheruser-generatedcontents.Aspecificformofopinionanalysisissentiment analysis, which is often used in opinionmining to identify sentiment, affect,subjectivity,andotheremotionalstatesinonlinetexts(Chenetal,2012:1176).Sentimentanalysiscanalsobeconsideredaspecificstepinopinionanalysis.Inourdiscussionwewillfirstfocusonopinionanalysisingeneral:wewillsubsequentlydiscusssentimentanalysisasaspecificstepinopinionanalysis.WealsorefertoourdiscussionondigitalsentimentindicesinChapter2.1.

Executionofopinionanalysis

Toexecuteopinionanalysisthefollowingstepsarerequired.

1. Socialdatacollectionandsetupdatabases2. Featuregenerationandpruning3. Opinionwordextraction4. Assessmentoforientationofopinionwords(sentiment)5. Summarygeneration(usingvisualization)

InFigure4.2.14wedisplay thesestepsof textanalyticsapproachesadapted fromHuandLiu(2004).Similarschemascanbefoundinotherstudiesaswell.

Figure4.2.14Textanalyticsapproach

Source:AdaptedfromHu&Liu(2004)

Socialdatacollectionanddatabasesetup

Asafirststeptheresearcherhastodecidewhattexttheywanttoanalyze.Thisstartswitha clear research objective. For example, if firms want to gain insights from customerreviewsofabrandorproduct, rawdataofreviewsshouldbeselected.However, if firmswanttomeasurethesentimentaboutaspecificbrand,dataonwhatisbeingwrittenaboutbrandsinmultiplesocialforashouldbecollected.Theseexamplesalreadyimplythatthecollectionofdataisnotanobviousstep.Thereareissuessurroundingwheretocollectdata,the included timeperiod,andwhichbrandsand/orproductsdatawillbecollected for. Ifthedataarebeingcollectedadatabasefortextdatashouldbesetup,whichcanbeusedforanalysis. To set up this database POS-tagging (part-of-speech tagging) is used, whichassigns or “tags” information to each word in a sentence. Tags mainly involve the

characteristicsof theword inasentence.Atagcouldbewhether theword isanoun,orwhetheritisaverboranadjective,etc.(seeexampleinFigure4.2.15).

Figure4.2.15IllustrationofPOStagging

This raw database is usually pre-processed,whichmight include the removal of stopwords,stemming,andfuzzymatching.Fuzzymatchingisusedtodealwithwordvariantsandmisspellings (Hu& Liu, 2004).An outcome of this step can be a “word cloud” (seeFigure4.2.16).

Featuregenerationandpruning

The second step involves the identification of relevant product or brand featureswhichpeoplehaveexpressedintheiropinions.Sometimesthiscanberathereasy.Forexample,ifareviewforadigitalcamerainvolvesasentencelike:“Thesizeofthecameraisgood,”itisclearthatthefeature“camerasize”ismeant.However, if thesentencewas:“Thiscamerawouldnoteasilyfitinyourpocket,”itismuchmoredifficultfortextanalyticstopickupthat this customer is also hinting at camera size. Explicit features are thus easy to find,whereas more implicit mentions of features are more difficult to find. Text analyticalpackageshavetendedtofocusonexplicitfeatures.Researcherswouldprobablymainlybeinterested in themost frequentlymentioned features.Anotherproblem is that customersmayusedifferentwords for similar features (e.g. sizeand largeness).Associativeminingcanbeusedtofindafrequent itemsset,whichisasetofwords(oraphrase)thatoccurtogetherinsomesentences(e.g.Hu&Liu,2004).Thisapproachwillnotleadtoaperfectsetoffeatures.Toachieveafinalandhopefullygoodsetoffeatures,pruningtechniquesareused.HuandLiu(2004)describetheuseoftwopruningtechniques:compactnesspruningandredundancypruning.Compactnesspruningchecks featureswith twoormorewordsandremovesthosefeaturesthatarelikelytobemeaningless.Redundancypruningfocuseson removing redundant features. For example, for mobile phones “life” is not a usefulfeature,whereas“batterylife”couldbemeaningful.Thisstepcanbedoneusingcomputerprograms.However,inmanycasesahumantextanalystisstillrequiredtoexecutepartsofthesesteps.

Figure4.2.16Illustrationofawordcloud

Oneimportantissueisthefactthataswellasfrequentfeatures,infrequentfeaturesarealso mentioned. It would be easy to ignore these infrequent features. However, thesefeatures can be interesting for a specific customer segment. And perhaps these featuresmayindicatesomeimportant,notwidelyspreaddevelopmentsinthemarket,whichmaybecomedominant.Soitmightbedangeroustoignorethem.Onewaytofindthesefiguresistomakematchesbetweeninfrequentfeaturesandfrequentlyusedopinionwords.

Opinionwordextraction

Inthenextstep,wordsinthesentencesthatprovideacustomersopinionaboutaproductfeatureorbrandareidentified.Usually,opinionsinsentencesaregiventhroughtheuseofadjectives(e.g.,nice,bad,etc.).Asentencecanbeconsideredanopinionsentencewhenitincludes adjectives. Researchers could decidewhether they include only those sentencesthatarelinkedtotheproductfeaturesorbrandsbeingstudied.Typicallyonewouldliketoinclude only sentences in the analysis in which opinion words are linked to productfeaturesorbrands.

Assessmentoforientationofopinionwords

Whentheopinionwordshavebeenidentified,itisimportanttounderstandtheirsemanticorientation.Thisbasicallymeanswhethertheworddeviatesfromthenormforitssemanticgroup.Andonecanalsodistinguishbetweenwordshavingadesirable state thathaveapositiveorientation,andwordshavingundesirablestatethathaveanegativeorientation.Examplesofwordshavingapositiveorientationareawesome,good,fantastic,etc.Wordshavinganegativeorientationcouldbesad,disappointed,awkward,etc.Somewordscanalso be neutral. In marketing research this orientation is frequently referred to as thevalence.Wordscanhaveanegativevalence,aneutralvalenceorapositivevalence.Othershavereferredtothisasthe“affectivecontent”(e.g.Ludwigetal.,2013).Theassessmentofthis valence or affective state is typically done in (brand) sentiment analysis and is thebasisforcalculatingdigitalsentimentindices(seeChapter2.1).Basedon thisorientation,

textanalyticscanalsoleadtosomequantitativeoutcomes.Thewordsinasentencecouldgetascore(e.g.−1negative,0neutral,1positive).Onecouldalsoseekwaystoassessthevalence strength. For example, while fantastic and good both have a positive valence,fantasticisconsideredasmuchmorepositivethangood.Inaquantitativeway“fantastic”couldgetascoreof+2and“good”ascoreof+1.Standarddictionariescanbeusedtoassessthe valence. For example, researchers have used the dictionary of affect in language asdeveloped by Sweeney and Whissel (1984)—see for example Verhoef, Antonides, & DeHoog(2004)—butalsootherdictionariescanbeused—seeforexampleLudwigetal.(2013).

Summarygeneration

Thefinalstepistocreateasummaryoftheresults.Thiscanbeasimplefrequencytable.For example, one could report the frequency of each mentioned feature. Subsequently,thesefeaturescanbelinkedtoopinions.Heretheoccurringfrequencyofnegative,neutraland positive valence can be reported. These descriptive analyses are the norm in textanalytics in marketing. Further analyses can, however, also be executed. For example,linkedbrandfeaturesorassociationscanbeusedtocreateavisualbrandassociationmap(Gensler,Völckner, Egger, Fischbach,& Schoder 2015; see also Figure4.2.17). Input fromtextanalytics(i.e.thevalencescores)canalsobeusedasinputinpredicting,forexample,brandsales,brandacquisition,andothermetrics (e.g.DeVries,2015;Srinivasan,Rutz,&Pauwels,2015).

Bigdataarea7:Socialnetworkanalysis

In today’s digital environment social networks have become prominent as a result of agrowing number of social communities: Facebook, LinkedIn, and in China, Renren, areprobably the most prominent more general communities. Beyond these generalcommunities, therearealsomanysmaller specializedcommunities that focuson specificthemessuchasmusic,sports,cooking,andlifestyle.Inmanyappshybridformsofservicesarecombinedwithsocialcommunities.Forexample, the runningandcyclingappStravahas services on running and cycling performance, but users are also connected to otherusers.

Figure4.2.17Numberoftweetsbytimeandsentiment

Socialnetworksareofcoursenotanewphenomenon.Theyhavebeenaroundforyears;for example almost twenty years ago Achrol and Kotler (1999) mentioned the growingimportance of networks in marketing and specifically business-to-business marketing.Withinsociologytheimportanceofsocialnetworkshasalsobeenstressed(e.g.Burt1987).Theincreasingpresenceofsocialcommunitieshasstirredupahugeamountofattentionfornetworks:wearenowconsideredtoliveintheeraoftheconnectedcustomer(Wuyts,Dekimpe, Gijsbrechts, & Pieters, 2010). As a consequence, customer-to-customer (C2C)interactions have become very important inmarketing, and it is generally accepted thatcustomersinfluenceeachother.Thisideaisnotnew.Inhisfamousdiffusionmodel,Bass(1969) introduced the notion of social contagion capturing this social influence throughmultiplemechanisms (Van denBulte& Stremersch, 2004). Social communities have alsoincreasedtheavailabilityofnetworkdata,whichinturnhasledtomorestudiesonsocialnetworkdatawithinmarketing.

Analyzingsocialnetworks

Weconsiderthefollowingstepsinnetworkdataanalytics:

1. Creationofnetworkdata

2. Descriptionofthenetworkusingnetworkmetrics3. Socialnetworktargetingandvaluation.

Socialnetworkdata5

Theway the network is defined is an important issue. There are three generalways tocollectnetworkdata:

1. Self-reportsonnetworks2. Zipcodelevelasaproxyforanetwork3. Useofactualnetworksusingdigitalandmobiletechnologies.

Thusfar,moststudieshaveusedeitherself-reporteddataorgeographicaldata(Zipcodes)tobuildanetwork.Although theuseof surveys is common (Wasserman&Faust, 1994),theyhaveseveraldrawbacks,includingdependenceonrespondents’memories,differencesacrossrespondents’interpretations,andself-reportbiases(Bertrand&Mullainathan,2001),allofwhichcanleadtoerroneousdescriptionsofthenetwork.Geographicaldataarealsocommonly used because most customer and marketing databases contain Zip codeinformation(Bell&Song,2007;Nam,Manchanda,&Chintagunta,2010).Theideaisthatifindividualsliveclosertogethertheycanbeinasocialnetwork.Inhisinfluentialworkonthe prescription of pharmaceutical drugs, Wharton professor Christophe van den Bulteused Zip-code data to create networks of physicians whowere likely to influence eachother (e.g. Iyengar, Van den Bulte, & Valente, 2011). However, social interaction is notmeasureddirectly,andsotheuseofthesedatarequirestheassumptionthatpeoplelivingclosetoeachotherinfluenceeachother(Choi,Hui,&Bell,2010;Iyengar,VandenBulte,&Choi,2011).Itisunlikelytobeanaccuratedescriptionsincesocialinteractionisbecominglessandlessdependentonspatialproximity(Haenlein,2011).

Asanalternative,onlinesocialnetworksareapotentiallyrichsourceofnetworkdata.Thenumberofpeopleinthesenetworksistypicallylarge,anddataarerelativelyeasytoobtain(Godes&Mayzlin,2004;Stephen&Galak,2010).Adisadvantageofthismethodisthedifficultyincombiningnetworkinformationandbehavioralandtransactionaldataforthe same person. Furthermore, people are typically connected in an online network tomanyotherswhoarenotrelevantfromasocialinfluenceperspectivesinceitrequiresonlyvery little effort to establish and maintain a ‘friendship’ tie (Ackland, 2009; Trusov,Bodapati,&Bucklin,2010).

Another alternative is the use of direct communication data from email or (mobile)telephony. In the latter case so-called call detail records (CDR) can be used to createnetworks. In CDR data, all phone calls and text messages are recorded individually. Aperson’s mobile phone network is a good proxy for his or her social network (Eagle,Pentland, & Lazer, 2009; Haythornthwaite, 2005) and has been used to model retention

(Nitzan&Libai,2011)andadoption(Hill,Provost,&Volinsky,2006;Risselada,Verhoef,&Bijmolt, 2014). The analysis of these networks is, however, not trivial because of thepotentialforprivacyinfringements(seeChapter3.3).

Socialnetworkmetrics

Social networks inmarketing aremainly used to identify those networks or specificallythose individuals thathaveanoverly strong influenceonother customers, and thus canhelp to diffuse ideas, opinions, and new products in the market. Social marketingcampaignssuchasviralmarketingandreferralcampaignsaimtofocusonthesecustomers.These influentialcustomersare typically the firstcustomersapproachedby the firminasocial campaign, andare considered the seedsof a campaign.These seedswill influenceothersandtherebyspreadthemessage(inthecaseofaviralcampaign)orconvinceotherstoact(inthecaseofareferralcampaign).

Animportantquestionishowtoidentifytheseinfluentialcustomers.Atraditionalwayhasalwaysbeen tousea surveymeasure:opinion leadership (e.g.Kratzer&Lettl,2009).Customers reporton theirbeliefson their influenceonother customers.Recent researchsuggests that thismeasure is actuallynot very good andhasnodirect impact on actualinfluence(Risselada,Verhoef,&Bijmolt,2015).There is thusadiscrepancybetweenself-reported influence andactual influence,which shouldnot comeas a surprise, as similardiscrepancies between other self-reportingmeasures and behavior have been found (e.g.betweenpurchaseintentionandpurchasebehavior).

Actualnetworkcharacteristicsdo,however,makeadifference.Thesenetworkmetricsallfocusonthecentralityofcustomersinanetwork,andspecificallythefollowingmetricsareproposed(VandenBulte&Wuyts,2007):

DegreecentralityBetweennesscentralityClosenesscentrality.

Degreecentrality(seeFigure4.2.18)measuresthenumberofcontactsanindividualhasinanetwork. Customers with a high degree centrality are called hubs, as through thesecustomersmanyothercustomerscanbereached (similar toairporthubs,where fromanairporthublikeAtlantaorAmsterdammanydestinationscanbereached).Oneconcernisthat these hubshavemany contacts, but their effort per contact could be lower. Severalstudieshave,however,shownthatthesehubsworkprettywell(e.g.Hinz,Skiera,Barrot,&Becker, 2011), and degree centrality predicts influence (Risselada et al., 2015). Degreecentrality only considers direct contacts, and does not consider the position of thecustomersinamoreextendednetwork.

Figure4.2.18Degreecentrality

Figure4.2.19Betweennesscentralityandclosenesscentrality6

Notes:aBetweennesscentralityforE=6bClosenesscentralityforE=1.2

The other twometrics have amore extended network perspective (see Figure 4.2.19).Betweennesscentralitymeasurestheextenttowhichanindividualisintheshortestrouteorpathbetweencustomersinanetwork.Theunderlyingideaisthatcustomerswithalowbetweennesscentralityhavemorepowerinanetwork.Closenesscentralitymeasurestheaveragenumberofstepsinanetworkittakestoreachallothercustomersinanetwork.Itbasicallymeasureshoweasyitisforcustomerstoreachallothercustomersinanetwork.A low closeness centrality suggests that this should be more easy.While these metricsprovideusefulinformationandindeedmatter(e.g.Katona,Zubcsek,&Sarvary,2011),theyareusuallymoredifficulttomeasure,asacompletenetworkisrequired.Degreecentralitydoesnotrequireafullnetworkstructure,asoneonlyhastocountthenumberofcontacts.

Aswellaslookingatnetworkcharacteristics,contactsbetweenindividualcustomersinanetworkcanalsobecharacterized.Twokeyvariablesaretiestrengthandhomophily.Tiestrengthreflectstheintensityofthecontactbetweencustomers.Whenatie isstrongtheintensityofcommunicationisstrongerandthusitislikelythatcustomershaveastrongerinfluence on each other (Nitzan&Libai, 2011). It can bemeasured by, for example, thenumber of contacts (e.g. calls,Whatsappmessages) or the distance between individuals.Homophilyreflectsthedegreetowhichcustomersinanetworkaresimilar.Itisgenerallyacceptedthatsimilarindividualstendtobehaveinasimilarway,becausecustomerstendtoassociatemorewithsimilarpeople(McPherson,Smith-Lovin,&Cook,2001).Onewayofmeasuring homophily is by scoring two individuals on whether they have similar

characteristics,suchasage,education,genderetc.(e.g.Risseladaetal.,2014).

Socialtargetingandnetworkvaluation

Inmany cases firmswill be happywith being able to identify influential customers.Asnotedearlier,thesecustomerscanthenbeusedasseedsinsocialcampaigns.Inanonlineenvironmentthisapproachisreferredtoasaviralmarketingcampaign.Subsequentlythesecampaignshavetobeevaluatedandfirmsmightbe interested inmeasuringtheeffectofthesecampaignsandtoassess theresultingnetworkvalueofcustomers.Theresponseofcustomers to viral marketing campaigns can be modeled. Van der Lans, Van Bruggen,Eliashberg&Wierenga,(2010)developedaviralbranchingmodel.Hinzetal.(2011)reportthat seeding indeedworks and createsmorevalue.As anext step customer engagementmetricssuchascustomerreferralvalue(CRV)andcustomerinfluencevalue(CIV)canbecalculated(Kumaretal.,2010;seealsoChapter2.2).

Network variables for each customer can also be used to predict behavior. Severalstudieshaveshownthatsimilarcustomerswithstrongtiesinfluenceoneanother.Soifafirmknowsthatacustomerhasdefected,theycouldidentifycustomersinthenetworkofthatcustomerwhoarealsoatriskfordefecting(Nitzan&Libai,2011). Ifacustomerhasadoptedanewproduct,individualsinhis/hernetworkcanbeidentifiedandcontactedtoinducenewproductadoptionaswell.Wewarn,however,thatthesocialinfluenceeffectsare less simple than usually considered, as social influence not only operates at thecustomerlevelbutalsoatthemarketlevelthrough,forexample,socialnorms(Risseladaetal.,2014).

Emergingtechniques

As a result of digital developments new techniques to analyze data will emerge. Textanalytics have improved considerably over the last decade. Nowadays, customersconstantlyuploadphotosandvideosusing,forexample,mobilephones,andthisnewdatacan also provide information about brands and customer experiences. Photo and videoanalyticswillemergetohelpgetrelevantmarketinginformationfromthesenewsourcesofdata.Establishedfirms,suchasMicrosoft,butalsonewfirms,aredevelopingtechniquestoanalyzevideos.7

Advances inmobilephoneusagewill also inducemoreattention formobileanalytics.Mobilescanbeusedto followcustomers instores, forexample.Mobileanalyticsprovidetechnologicalsolutionsforretailersbydevelopingaggregatereportsusedtoreducewaitingtimesatcheckouts,optimizestorelayouts,andtounderstandconsumershoppingpatterns.The reports are generated by recognizing theWi-Fi or BluetoothMACaddresses of cellphonesastheyinteractwithstoreWi-Finetworks.Mobiledatacanprobablybecomeverypowerful.Sofarknowledgeofthisislimited.However,infiveyearsfromnowthesituationmighthavedramaticallychanged,asfirmsarestartingtoexperimentwiththesedataandmanystudieswillbeexecutedusingthesedata.

Conclusions

Bigdatahavean impactonanalyticsasnewdatabecameavailableandnewapplicationareasbecameprominent. In this in-depthchapterwehavediscussed sevennewbigdataareas,someofwhichhavebeenthereforawhilealready(e.g.webanalytics),whileothersarerathernew(e.g.attributionmodeling,sociallistening).Wehaveaimedtodiscusseachoftheseareasinsufficientdepthtoenablethereadertounderstandthebasicsoftheareaandthemanyavailabletechniques.Wenotethatsomeofthetechniquesbuildonexistingtraditional techniques. For example, in the case of attributionmodeling logitmodels areproposedasawaytostudythecontributionofeachtouchpointtoconversion.Finally,wediscussed some emerging areas, which are yet not sufficiently developed to extensivelydiscuss in this book. In five years from now this situation might, however, be totallydifferent.

Notes

1ThissectionbenefitedstronglyfromtheexcellentoverviewprovidedbyChung&Wedel(2014)onpersonalizationin

services.

2Seewww.google.org/flutrends/about/how.html(accessedSeptember23,2015).

3Seehttp://bits.blogs.nytimes.com/2014/03/28/google-flu-trends-the-limits-of-big-data/?_r=0).

4Seehttp://gking.harvard.edu/files/gking/files/0314policyforumff.pdf (accessedSeptember22,2015)Thisarticlemakes

veryinsightfulreadingonproblemswithbigdata.

5ThissectionisbasedonBeckers,RisseladaandVerhoef(2014:97–120).

6WethankHansRisseladaforsharingthesefigures.

7Seehttp://techcrunch.com/2015/05/04/video-to-data/(accessedSeptember24,2015).

References

Achrol,R.S.,&Kotler,P.(1999).Marketinginthenetworkeconomy.JournalofMarketing,63(4),146–163.

Ackland,R.(2009).Socialnetworkservicesasdatasourcesandplatformsfore-researchingsocialnetworks.SocialScienceComputerReview,27(4),481–492.

Allenby,G.M.,Rossi,P.E.,&McCulloch,R.E.(2005).HierarchicalBayesmodels:apractitionersguide.SSRWorkingPaperSeries.RetrievedfromSocialScienceResearchNetwork.RetrievedOctober1,2015fromhttp://ssrn.com/abstract=655541.

Ansari,A.,Essegaier,S.,&Kohli,R. (2000). Internetrecommendationsystems.JournalofMarketingResearch,37(3),363–375.

Ansari, A.,Mela, C. F., & Neslin, S. A. (2008). Customer channel migration. Journal ofMarketingResearch,45(1),60–76.

Arora,N.,Dreze,X.,Ghose,A.,Hess,J.D.,Iyengar,R.,Jing,B.,Joshi,Y.,Kumar,V.,Lurie,N.,Neslin,S.,Sajeesh,S.,Su,M.,Syam,N.,Thomas,J.,&Zhang,Z.J. (2008).Puttingone-to-onemarketing towork:Personalization, customization, and choice.MarketingLetters,19(3–4),305–321.

Ataman,M.B.,Mela,C.F.,&VanHeerde,H.J.(2008).Buildingbrands.MarketingScience,27(6),1036–1054.

Bass, F. M. (1969). A new product growth for model consumer durables.Managementscience,15(5).215–227.

Beckers, S. F. M., Risselada, H., & Verhoef, P. C. (2014). Customer engagement: a newfrontierincustomervaluemanagement.InR.T.Rust,&M.H.Huang,(Eds),HandbookofServiceMarketingResearch(pp.97–120).CheltenhamUK:EdwardElgarPub.

Bell,D.R.,&Song,S.(2007).NeighborhoodeffectsandtrialontheInternet:Evidencefromonlinegroceryretailing.QuantitativeMarketingandEconomics,5(4),361–400.

Bertrand,M.,&Mullainathan,S.(2001).Dopeoplemeanwhattheysay?Implicationsforsubjectivesurveydata.TheAmericanEconomicReview,91(2),67–72.

Bitner, M. J., Ostrom, A. L., & Morgan, F. N. (2008). Service blueprinting: A practicaltechniqueforserviceinnovation.CaliforniaManagementReview,50(3),66–94.

Bodapati,A.V.(2008).Recommendationsystemswithpurchasedata.JournalofMarketingResearch,45(1),77–93.

Bucklin,R.E.,&Sismeiro,C.(2003).Amodelofwebsitebrowsingbehaviorestimatedonclickstreamdata.JournalofMarketingResearch,40(3),249–267.

Bucklin, R. E., & Sismeiro, C. (2009). Click here for internet insight: Advances in

clickstreamdataanalysisinmarketing.JournalofInteractiveMarketing,23(1),35–48.

Bucklin,R.E.,Lattin,J.M.,Ansari,A.,Gupta,S.,Bell,D.,Coupey,E.,Little,J.D.C.,Mela,C.,Montgomery,A.,&Steckel,J.(2002).ChoiceandtheInternet:Fromclickstreamtoresearchstream.MarketingLetters,13(3),245–258.

Bult, J. R., &Wansbeek, T. (1995). Optimal selection for directmail.Marketing Science,14(4),378–394.

Burt,R.S.(1987).Socialcontagionandinnovation:Cohesionversusstructuralequivalence.AmericanJournalofSociology,92(6),1287–1335.

Chen,H.,Chiang,R.H.L.,&Storey,V.C.(2012).Businessintelligenceandanalytics:Frombigdatatobigimpact.MISQuarterly,36(4),1165–1188.

Choi, J., Hui, S. K., & Bell, D. R. (2010). Spatiotemporal analysis of imitation behavioracrossnewbuyersatanonlinegroceryretailer.JournalofMarketingResearch,47(1),75–89.

Chung,T.S.,&Wedel,M.(2014).Adaptivepersonalizationofmobileinformationservices.InR.T.Rust,&M.H.Huang,(Eds),HandbookofServiceMarketingResearch(pp.395–412).Cheltenham:EdwardElgarPub.

Chung,T.S.,Rust,R.T.,&Wedel,M.(2009).Mymobilemusic:Anadaptivepersonalizationsystemfordigitalaudioplayers.MarketingScience,28(1),52–68.

DavidShepardAssociates. (1999).Thenewdirectmarketing:How to implementaprofit-drivendatabasemarketingstrategy.Europe:Mcgraw-HillEducation.

DeHaan,E.,Verhoef,P.C.,&Wiesel,T.(2015).Thepredictiveabilityofdifferentcustomerfeedbackmetrics for retention. International Journal ofResearch inMarketing,32(2),195–206.

DeHaan,E.,Wiesel,T.,&Pauwels,K.(2013).Whichadvertisingformsmakeadifferenceinonlinepathtopurchase.MSIWorkingPaperSeries,13(104),Boston.

DeVries,L.(2015).ImpactofSocialMediaonConsumersandFirms.Doctoraldissertation,UniversityofGroningen.

Eagle,N.,Pentland,A.S.,&Lazer,D.L.(2009).Inferringfriendshipnetworkstructurebyusingmobile phone data. Proceedings of the National Academy of Sciences, 106(36),15,274–15,278.

Edwards,G.(2014).Howtoembraceretail’snewesttrend:Webrooming.RetrievedfromRetailTouchPointsBlog.RetrievedSeptember11,2015fromhttp://retailtouchpoints.tumblr.com/post/59107311088/how-to-embrace-retails-newest-trend-webrooming

Gensler, S., Verhoef, P. C., & Böhm,M. (2012). Understanding consumers’ multichannel

choicesacrossthedifferentstagesofthebuyingprocess.MarketingLetters,23(4),987–1003.

Gensler, S., Völckner, F., Egger, M., Fischbach, K., & Schoder, D. (2015). Listen to yourcustomers: insights into brand image using online consumer-generated productreviews.InternationalJournalofElectronicCommerce,20(1),112-–41.

Godes, D., & Mayzlin, D. (2004). Using online conversations to study word-of-mouthcommunication.MarketingScience,23(4),545–560.

Haenlein, M. (2011). A social network analysis of customer-level revenue distribution.MarketingLetters,22(1),15–29.

Hamilton,J.D.(2008).Regimeswitchingmodels.InDurlaufS.&BlumeL.,(Eds),PalgraveDictionaryofEconomics.London:PalgraveMcMillan.

Haythornthwaite,C.(2005).SocialnetworksandInternetconnectivityeffects.InformationCommunication&Society,8(2),125–147.

Hill,S.,Provost,F.,&andVolinsky,C.(2006).Network-basedmarketing:identifyinglikelyadoptersviaconsumernetworks.StatisticalScience,21(2),256–276.

Hinz,O.,Skiera,B.,Barrot,C.,&Becker,J.U.(2011).Seedingstrategiesforviralmarketing:Anempiricalcomparison.JournalofMarketing,75(6),55–71.

Hu, M., & Liu, B. (2004). Mining opinion features in customer reviews. AmericanAssociationforArtificialIntelligence,755–760.

Iyengar, R., Van den Bulte, C., & Choi, J. (2011). Distinguishing among multiplemechanismsofsocialcontagion:Social learningversusnormative legitimation innewproductadoption.Workingpaper.TheWhartonSchool,UniversityofPennsylvania.

Iyengar,R.,VandenBulte,C.,&Valente,T.W. (2011).Rejoinder–furtherreflectionsonstudyingsocialinfluenceinnewproductdiffusion.MarketingScience,30(2),230–232.

Johnson,E.J.,Bellman,S.,&Lohse,G.L. (2003).Cognitivelock-inandthepowerlawofpractice.JournalofMarketing,67(2),62–75.

Katona,Z.,Zubcsek,P.P.,&Sarvary,M.(2011).Networkeffectsandpersonalinfluences:Thediffusionofanonlinesocialnetwork.JournalofMarketingResearch,48(3),425–443.

Konuş,U.,Neslin,S.A.,&Verhoef,P.C.(2014).Theeffectofsearchchanneleliminationonpurchaseincidence,ordersizeandchannelchoice.InternationalJournalofResearchinMarketing,3(1),49–64.

Konuş,U.,Verhoef,P.C.,&Neslin,S.A.(2008).Multiannelshoppersegmentsandtheirantecedents.JournalofRetailing,84(4),398–413.

Kratzer,J.,&Lettl,C.(2009).Distinctiverolesofleadusersandopinionleadersinthesocial

networksofschoolchildren.JournalofConsumerResearch,36(4),646–659.

Kumar, V., Aksoy, L., Donkers, B., Venkatesan, R., Wiesel, T., & Tillmanns, S. (2010).Undervalued or overvalued customers: Capturing total customer engagement value.JournalofServiceResearch,13(3),297–310.

Lazer,D.,Kennedy,R.,King,G.,&Vespignani,A.(2014).TheparableofGoogleflu:Trapsinbigdataanalysis.Science,343(6176),1203–1205.

Ludwig,S.,DeRuyter,K.,Friedman,M.,Brüggen,E.C.,Wetzels,M.,&Pfann,G. (2013).More thanwords: The influence of affective content and linguistic stylematches inonlinereviewsonconversionrates.JournalofMarketing,77(1),87–103.

Macdonald,E.K.,Wilson,H.N.,&Konuş,U.(2012).Bettercustomerinsight–inrealtime.HarvardBusinessReview,90(9),102–108.

Manning,C.,&Schütze,H.(1999).Foundationsofstatisticalnaturallanguageprocessing.Cambridge:MITPress.

McPherson,M., Smith-Lovin, L.,&Cook, J.M. (2001). Birds of a feather:Homophily insocialnetworks.AnnualReviewofSociology,27(1),415–444.

Melis,K.,Campo,K.,Breugelmans,E.,&Lamey,L.(2015).Theimpactofthemulti-channelretailmixononlinestorechoice:Doesonlineexperiencematter?JournalofRetailing,9(2),272–288.

Mincer,J.(1969).Economicforecastsandexpectations:Analysisofforecastingbehaviorandperformance.NewYork:NationalBureauofEconomicResearch.

Nam, S., Manchanda, P., & Chintagunta, P. K. (2010). The effect of signal quality andcontiguous word of mouth on customer acquisition for a video-on demand service.MarketingScience,29(4),690–700.

Neslin,S.A.,Grewal,D.,Leghorn,R.,Shankar,V.,Teeling,M.L.,Thomas,J.S.,&Verhoef,P. C. (2006). Challenges and opportunities in multichannel customer management.JournalofServiceResearch,9(2),95–112.

Nitzan, I.,&Libai,B. (2011). Social effects on customer retention. Journal ofMarketing,75(6),24–38.

Rapp,A.,Baker,T.L.,Bachrach,D.G.,Ogilvie,J.,&Beitelspacher,L.S. (2015).Perceivedcustomershowroomingbehaviorandtheeffectonretailsalespersonself-efficacyandperformance.JournalofRetailing,91(2),358–369.

Risselada,H.,Verhoef,P.C.,&Bijmolt,T.H.A.(2014).Dynamiceffectsofsocialinfluenceand direct marketing on the adoption of high-technology products. Journal ofMarketing,78(2),52–68.

Risselada,H.,Verhoef,P.C.,&Bijmolt,T.H.A.(2015).Indicatorsofopinionleadershipin

customernetworks:Selfreportsanddegreecentrality.MarketingLetters.Forthcoming.

Rust,R.T.,&Chung,T.S.(2006).Marketingmodelsofserviceandrelationships.MarketingScience,25(6),560–580.

Rust, R. T., & Verhoef, P. C. (2005). Optimizing the marketing interventions mix inintermediate-termCRM.MarketingScience,24(3),477–489.

Salton,G.(1989).Automatictextprocessing:thetransformation,analysis,andretrievalofinformationbycomputer.Boston,USA:Addison-WesleyLongmanPublishing.

Shi,S.W.,&Zhang,J.(2014).Usageexperiencewithdecisionaidsandevolutionofonlinepurchasebehavior.MarketingScience,33(6),871–882.

Snijders,T.A.B.,&Bosker,R.J. (2012).Multilevelanalysis:Anintroductiontobasicandadvancedmultilevelmodeling.London:SagePublishers.

Srinivasan,S.,Rutz,O.,&Pauwels,K. (2015).Paths toandoffpurchase:Quantifying theimpactoftraditionalmarketingandonlinecustomeractivity.Journalof theAcademyofMarketingScience.Forthcoming.

Stephen,A.T.,&Galak,J.(2010).Thecomplementaryrolesoftraditionalandsocialmediaindrivingmarketingperformance.Workingpaper,CarnegieMellon.

Sweeney,K.,&Whissel,C.(1984).Adictionaryofaffectinlanguage:I.Establishmentandpreliminaryvalidation.PerceptualandMotorSkills,59,695–698.

Trusov, M., Bodapati, A. V., & Bucklin, R. E. (2010). Determining influential users inInternetsocialnetworks.JournalofMarketingResearch,47(4),643–658.

VandenBulte,C.,&Stremersch,S. (2004).Socialcontagionand incomeheterogeneity innewproductdiffusion:Ameta-analytictest.MarketingScience,23(4),530–544.

Van denBulte,C.,&Wuyts, S.H.K. (2007).Social networks inmarketing. Boston:MSIRelevantKnowledgeSeries.

VanderHeijden,M.,DeHaan,E.,&Hoving-Wesselius,T.(2014).Eenmodelmatigeaanpakomheteffectvanonlineadverterenopconversieteachterhalen.InA.E.Bronner(Ed.),JaarboekMarktOnderzoekAssociatie.Haarlem:SpaarenHout.

VanderLans,R.,VanBruggen,G.,Eliashberg, J.,&Wierenga,B. (2010).Viralbranchingmodelforpredictingthespreadofelectronicwordofmouth.MarketingScience,29(2),348–365.

Verhoef,P.C.,Antonides,G.,&DeHoog,A.N.(2004).Serviceencountersasasequenceofevents:Theimportanceofpeakexperiences.JournalofServiceResearch,7(1),53–64.

Verhoef, P.C.,Kannan, P.K.,& Inman, J. (2015). Frommulti-channel retailing to omni-channelretailing: Introductiontothespecial issueonmulti-channelretailing.JournalofRetailing,9(2),174–181.

Verhoef,P.C.,Neslin, S.A.,&Vroomen,B. (2007).Multichannel customermanagement:Understandingtheresearch-shopperphenomenon.InternationalJournalofResearchinMarketing,24(2),129–148.

Wasserman,S.,&Faust,K.(1994).Socialnetworkanalysis–Methodsandapplications.UK:CambridgeUniversityPress.

Wedel, M., & Kamakura, W. A. (2000). Market segmentation: conceptual andmethodologicalfoundations.InInternationalseriesinquantitativemarketing.Boston:Kluwer.

Wuyts, S. H. K., Dekimpe, M. G., Gijsbrechts, E., & Pieters, R. (2010). The connectedcustomer: The changing nature of consumers and business markets. New York:RoutledgeAcademic.

Zhang, J.,&Wedel,M. (2009).Theeffectivenessof customizedpromotions inonlineandofflinestores.JournalofMarketingResearch,46(2),190–206.

4.3Creatingimpactwithstorytellingandvisualization1

Introduction

In thepreceding in-depth chapterswehavediscussed analytical techniques. Inorder foranalyststocreateimpactwithbothtraditionalandbigdataanalytics,howtheanalyticalresultsarecommunicatedisessential.Oneofthemaindangersofanalysisisthatareportis not even presented, or ends up in the never-opened desks ofmanagers and thereforeneverhasanyeffectonmanagementdecisions.Tohave impact two issuesareof crucialimportance:

1. The presence of a clear storyline in which the message of the implications isconciselydiscussed

2. The use of powerful visualization of the analytical results (i.e. effective use ofvisualaids).

Theimportanceofthesetwoissuesisincreasinglypresent.Thegrowthintheavailabilityofcontinuousdigital informationresultsinpeoplehavinglesstimeavailabletoattendtocommunications.Consumersaswellasmanagers (beingalsoconsumers) switchbetweenmultiple devices continuously to readmessages (e.g. onWhatsApp), emails, news apps,socialmedia,etc.Itisnowverycommonthatinmeetingsparticipantsdonotgivetheirfullattentiontopresentationsbecauseofbeingdistractedbywhatelseisbeingshownontheirtabletsormobilephones. It is thereforeveryimportantthatthepresentationsofresearchresults are sufficiently clear and attract attention. Further, today’s overload of digitalinformationmeansthatmanagershavetofindwaysoffilteringtherightinformationandinterpretingtheresults.Thisinformationloadisnotnew.Overthelasttwodecadespeoplehavebeenaddressingthegrowingimportanceofinformationstress.Thisarisesbecauseofthegrowingdivergencebetweenwhat information isavailableandwhatwecanprocess(seeFigure4.3.1).Itcanbeviewedasablackholebetweendataandknowledgethatstartstoexistwheninformationisnottellinguswhatwewantandshouldknow.Foralongtime,managersdidnotunderstandthattheydidnotknow.However,nowtheyunderstandwhattheydonotknowandasaconsequencetheyfeelinformationstress(Wurman,1989).

Thisincreasingoverloadiscallingforsolutions.Oneofthesolutionshasbeentoworkwithinfographics.Theseinfographicsareatypicalwaytomakecomplexinformationmoreaccessible for the reader. They exist inmany forms. Itmight, however, be questionablewhetherinfographicsareeffective.Aninfographicusuallyinvolveslittlestructureandtellsmany facts in the form of text and figures. The graphs or pictures are mainly used asillustrationstomaketheinformationmoreattractive.

Figure4.3.1Informationoverload

We strongly believe in the combination of data, storytelling, and visualization. Thesethreeelementsshouldstrengtheneachotherinsuchawaythattheinformationhasimpact(seeFigure4.3.2).Ifanalysts,basedonthestrengthsoftheirdataandanalytics,areabletotellastrongstoryandprovidestrongvisualizations,theyshouldhaveastrongimpact.Akindof“sweetspot”isachieved,asstrongvisualizationsandgoodstorytellingcombinedwith excellent data and analytics will be well received by managers in an era ofinformationoverload.Toachievethismulti-disciplinaryskillsarerequired.Thisisnoteasy,as frequently the analyst focusesonnumbers andmaybeunskilled at communicatingastrongstory.WewilldiscussthisgeneralissueinmoredetailinChapter5.

Figure4.3.2Sweetspotofdata,storyandvisual

Inthisin-depthchapterwediscusshowtobuildupagoodstorylineinanalyticalreportsandpresentationsandhowstrongvisualizationcanbeachieved.Beforedoingso,wefirstfocusonwhymanyanalyticalprojectswithstrongdataandanalyticsfailtohaveimpact.

Failurefactorsforcreatingimpact

Manyanalystshaveprobablyexperiencedcarryingoutaverynice study,but in theendtheir work did not change marketing strategies or tactics. Why does this occurs? It isdefinitelynot theanalyticalquality that iscausing theproblem.Thenumbershavebeencrunched in the right manner, the right research questions have been studied, but stillimpact is limited. This is probably the greatest frustration ofmany analysts.Having noimpactwill in the long run threaten the positon of the analytical functionwithin firms.Creating impact is also strongly required to create value with big data analytics! It istherefore very important to understand why reports do not have an impact and whattypicallygoeswrong.Wehavealreadyemphasizedthehugeimportanceofstorytellingandvisualizations. Based on our experience in analytical functions in many firms, we canidentifysomespecificissuesthatfrequentlygowrongandreducetheimpactofanalyticalexercises:

There is no structure to the report. One frequently reports independent analysesthat are not strongly related and as a consequencemanyunrelatedmessages arebeingcommunicatedinsteadofafewstronglyrelatedmessages.There are no strong and clear conclusions ormessages. The findings are nice toknow, but it is unclearwhat themanager should dodifferently after reading thereport.Oneeasilygetsthe“Sowhat?”response.The reports include toomanypagesor slides.Moreover, theconclusionsareonlyreportedat theendof thereport.Theattentionof thereaderhasdiedoutby theend of the report and conclusions and implications are not read or mentallyprocessed!Theconsequenceisnoimpact.The main findings are good and understood, but combined with nice-to-knowirrelevantinsights.Theseinsightsdistractthereaderandresultinlessfocusonthemainmessageofthestudy.Thereareinconsistenciesinthereport.Thiscreatesadiscussiononthecontentandmaycreateconfusion,reducingtheperceivedreliabilityoftheresults.Thereportfocusestoomuchonthestatisticaldetailsandreportingonwhyspecificmethodshavebeenchosen.Althoughthisishighlyvaluedinscientificpublications,managershavenostronginterestinthedetails.Insteadofreportingthisinthemaintext,itcanbeprovidedasanappendix.The slides (or pages) are packed withmanymessages and as a result look verycrowded.Assumingthatnormalhumanshaveonlylimitedprocessingcapacityandareeasilydistracted,theycanusuallyprocessonlyalimitednumberofpoints.Analystsfrequentlyonlyreportmanynumbersinsteadofgraphs.Numbersarelesseasilyprocessedthanvisualgraphs.And if graphs are being used, they are too complex and provide too much

information.As a consequence it becomes a puzzle formanagers to pick out therightinformation.

Alltheseissuesrelatetoweakcommunication.Improvedcommunicationcanhappenwhenanalystslearnhowtobuildastrongfocusedstoryfortheirresultsandareabletovisualizethemintherightway.

Storytelling

Oneofthebasicprinciplesofagoodreportorpresentationisthatithasacoremessage.This coremessage should be introduced with a specific situation and complication andshould subsequently be underpinned with arguments. This approach is based on thepyramid principle as advocated by Barbara Minto (2009).2 At the end of the 1960s sheworkedasaconsultantatMcKinsey&Company,whereshefocusedonthedevelopmentofmethods tohelp their advisors to structure their presentations and reports.Thepyramidreferstotheprinciplethateachadviceshouldhaveapyramid-likestructure.Atthetopofthepyramidistheadvice,andbelowthetopstructure,indifferentpointsorparagraphs,isthe motivation. If the motivation is divided into multiple subpoints or issues a newpyramidstartstoexist.Subsequently,oneprovidesapowerfuldiscussion(ordescription)ofthe complication to introduce the key-message. In our experience of giving advice tocompanies based on analytics, we have observed that this pyramid principle is verypowerful. It really strengthens the impact of analytics. Schematically this results in thestructureasdisplayedinFigure4.3.3.

Tounderstandwhy thismethodcanbepowerful,we first considerwhat is frequentlybeingdonewhenreporting.Normally,analystsstart todiscusswhattheyhavedoneinakind of chronical order. Theywant to show themanager their analytical road trip fromproblemstatementtoendresults.Theanalystsonlyendtheirpresentationorreportwiththe important conclusions.They alsowant to be as complete as possible and aim to telleverydetail.Thisisverylogical.Analystshavebeentrainedlikethisinuniversities.Whenwritingathesis,theystartwithaproblemstatement,discussthetheory,thedatacollection,the analytical method, the results section and end with important conclusions. Similarstructures can be found inmany scientific papers, and thismay potentially explain thelimited impact of scientific papers on practice (Roberts, Kayandé, & Stremersch, 2014).However,whendoing this, themanagerwith limited timeandattentiononlygets to themostimportantresultsattheendofthereportorwhenthesessionisalmostfinished.

Figure4.3.3Buildingblocksforaclearstoryline

A report has muchmore impact when its core message and the context are directlyunderstood. By “directly”wemean that this should be set out at the beginning of eachreportorpresentation.Thecoremessagecanthenbeunderpinnedwithalimitednumberof arguments—one frequently uses seven as a kind of rule of thumb, a kind ofmagicalnumber basedonMiller’s law.The cognitive psychologistGeorgeA.Miller of PrincetonUniversityhas shown that there are severe limits inour capacity toprocess information(Miller,1956).Hisworkhasbeeninterpretedtomeanthattheaveragenumberofobjectsanaverage human can hold in memory is around 7. This strongly suggests limiting thevolumeofmessagesandargumentsbeingdiscussed.

Theabovediscussion clearly shows that there is amismatchbetweenhowananalystpresents an analysis and how it should be presented. An analyst frequently solves theproblemwithabottom-uptypeofapproach.However,effectivecommunicationsuggestsatop-down approach (see Figure 4.3.4). It is essential for analysts to understand thisdifference. When finishing a project and preparing the report and/or presentation theyshouldgetoutof theanalyticalmodeandmovetotheeffectivecommunicationmindset.Wehaveobservedthatanalyststypicallyfindthisdifficult,giventhattheytendtofocusondetailsandfrequently forget theoverallpictureandwhytheanalysis isbeingdone. It isthereforeimportanttoworkinanalyticalteams,whereeffectivecommunicationskillsareembeddedintheteam(seealsoChapter5).

Checklistforaclearstoryline

Theabovediscussionprobablyseemsratherintuitive,buthowcanitsconclusionsactuallybe implemented?We take the schema as shown in Figure 4.3.3 as a starting point, andbrieflypointtosomeissuesrequiringattention.

Situation

Figure4.3.4Analysisprocessvs.effectivecommunication

Whendescribingtheinitialsituation,thefollowingissuesrequireconsideration:

Is the discussed situation not controversial? Does the description in itself raisespecificquestionsand/oradebate? If the latteroccurs itwillbemoredifficult todiscussthecoremessage.Does the audience recognize the described situation? If so, they will be morereceptive.Is the situation description underpinned with figures and are these figuresunderstood and believed in the organization? If the latter is not the case thesituationdescriptionwillbelesseffective.Still,figuresshowingspecificproblemsinperformance(e.g.decreaseinnetpromoterscore(NPS),orincreaseinchurnrates)areveryimportanttoshowtherelevanceofthereport.Does the situation description create a complication and a specific researchquestion?

Complication

The complication can be defined as describing the problemor challenge.This should bedirectlyrelatedtothesituationdescription.Therearesomespecificissuesheretoconsideraswell.

Does the complication describe its potential impact for the organization? Forexample,inthecaseofdecreasingchurnrates,theimpactcouldmeanlowersalesovertime,decreasingmarketshare,andlowerprofitability.Is thecomplicationfirmlyunderpinnedwithargumentsand/orfigures?Againweadvisefocusingonfiguresthatcanbedirectlylinkedtoperformanceconsequences.Thiswillcreateastrongerbeliefintherelevanceoftheexecutedstudy.

Message

Whendiscussingthemessagethefollowingissuesshouldbechecked:

Isthereasinglecoremessageoraretheremanymessages?Weprefertoworkwithasinglecoremessagetohavemoreimpactofthatsinglemessage.Does the core message create some curiosity or question? Curiosity will createattentionandadesiretolisten.Doesthecoremessageprovideananswertothecomplication?

Underpinningthemessage

Whenprovidingargumentsforthemessage,itisimportanttoassesswhat,why,andhow

the argument is being used.Does the argument reallymake sense andwill it provide astrongunderpinningofthemessage?Specificissuesthatrequireattentionare:

Do the arguments link with questions a reader will ask when reading the coremessage? It is thus very important to understand how managers will react andwhatquestionswillcomeupwhenthecoremessageisbeingread.Aretheargumentscompleteandmutuallyexclusive?Acompletelistofargumentswill show that the analyst has seriously thought about the provided conclusion.Mutuallyexclusiveargumentsmeansthatthereisnooverlapintheconclusionsorintheopportunitiesyoufound.Thereshouldnotbetoofewnortoomanyarguments.Ageneralruleofthumbisthatthereshouldbeaminimumoftwoargumentsandamaximumoffive.Oneshouldstartwiththemostimportantandconvincingargumentandendwiththeleastimportantone.Aretheargumentscompatible?Forexample,whenhavingstrategicarguments,oneshould not have arguments that aremore tactical.Or if arguments are based onfacts,itisprobablynotwisetousesentimentsaswell.

InFigure4.3.5wegivesomeexamplesofstorylinesthatdifferintheirpurpose.Inexampleoneweshowhowonecancomeupwithbusinessopportunitiesthatachievethebusinesstarget.

Visualization

Visualizationisofutmostimportanceincreatingimpactwithdataanalytics.Thereasonisactually rather simple: “Apicture iswortha thousandwords.”Theability tounderstandand extract value from data is much easier when it is done through data visualizationrather thanfromlookingat therawdataor thesimplestatisticsof thedata. In1973, thestatisticianFrancisAnscombedemonstratedtheimportanceofgraphingdata.Anscombe’sQuartet shows how four sets of datawith identical simple summary statistics can varyconsiderablywhengraphed(seeFigure4.3.6).

Figure4.3.5Examplesofdifferentstorylinesfordifferentpurposes

Figure4.3.6GraphofAnscombe’sQuartetdatatable

Source:AdaptedfromAnscombe(1973)

There are also some statistics underlying these claims. For example, if information istransferredorallyonly10%ofthereceiverscanrememberthatinformationafter72hours;this percentage rises to 65% if the information is visualized. This is called the “picturesuperiorityeffect”(Paivio&Csapo,1973).SoifyoulookatFigure4.3.7itismorelikelythatyouwillremembertherightpart(withtheapple)thantheleftpart.

Visualizationinanalyticsisusedformanypurposes.Generallytherearethreeobjectivesthatcanbeachievedbyvisualizingdata:

ExplorationofdataUnderstandandmakesenseofthedataCommunicatetheresultsoftheanalysis.

Thefirsttwoobjectivesaregenerallypartsoftheanalysisprocess.Beforerunningallkindsofanalysesitiswisetoexplorethedatawithvisualstohelpmakesenseofthem.Thiscanleadtoimmediatevaluableinsightsandtheunderstandingofpotentialrelationshipsinthedata. The last objective is clearly linked to the presentation of the results and creatingimpact.Ofcourseinsomecasesvisualizationofdataexplorationscanalsobeusedinthepresentationifitunderpinsthemainmessageofthereport.Visualizationsinmanyformscan be usedwherever results are presented, such as in reports, presentations,marketingdashboards,andwebsites.Wewillprobablyobservenewtrendsinwhichappswillbeusedtovisualizedata.

Figure4.3.7Thepicturesuperiorityeffect

Inthenextsectionsweaimtoprovidesomepracticalguidelinesonhowtoeffectivelyvisualizewhencommunicatingtheresults.Wespecificallyfocuson:

ChoosingtherightcharttypeDesignofthechart

We will also provide you with some practical tips and tricks to further improve thevisualization.

Choosingthecharttype

Acommonmistakewhenusingachartistojustchooseachart,forexampleabarchartora scatter plot, and assume that using such a chart in a report or presentation will besufficientanditwillbeself-explanatory.However,thisisfrequentlynotthecase.Choosingtherightchartformattocommunicatetheanalyticalresultsshouldbedonecarefully.Theproblemthat researchers face is that therearemanygraph types, styles,andmethods topresentdata.Thismakesitdifficulttochoosetherightformat.Tofindtherightcharttype,itisimportanttoknowthattherearefourcoretypestovisualizedata:

1. Showingarelationshipbetweendatapoints2. Comparingdatapoints3. Showingthecompositionofdata4. Showingthedistributionofdata.

Whenchoosingtherightcharttypeitisfirstimportanttoassesswhichgraphtypefitsbestwiththemessage,theoneaimtoconvey.Indoingso,onehastoconsiderthepurposeofthe graph. The above distinction in the variousways to visualize can be helpful in thisrespect.Wewillthereforediscussthedifferenttypesofdatavisualizationforeachofthesefourtypes.

Relationshipbetweendatapoints

Figure4.3.8Relationshipcharts

Agraphdisplayingarelationshipaimstoshowtheassociationorcorrelationbetweentwoormorevariablesthroughthedatapresented,suchasshowingtherelationshipbetweenin-storesalesandholidays.Themostcommon“relationcharts”thatdothisarescatterplotsandbubblecharts.Butyoucanalso thinkaboutgeographicorgeospatialgraphsorevennetworkchartswhenyouwanttoshowtherelationbetweenobjects(seeFigure4.3.8).

A scatter chart is used to show a relationship between two variables (X, Y) todetermine if they tend tomove in the same or opposite directions. An examplemightbeplottingNPS(X)andretention(Y)forasampleofmonths.A bubble chart is an extension of the scatter chart, adding a third variable. Thisendsupbeingreflectedinthesizeofthebubble.Forexample,whenshowingtherelationship betweenNPS and retention, the size of the bubblemight reflect the

numberofcustomersataspecificdatapoint.A geographic map typically shows the relationship between geographic locationandavariable.Forexample,onemightaimtoshowthesalesvolumeperregionorcountry.Geographicunitscanbecountries,regions,Zipcodeareas,etc.Anetworkcharttypicallyshowstherelationshipsbetweenobjects(seeFigure2.1.6)orindividuals.Thislasttypeofvisualizationcanbeusedinsocialnetworkanalysis,whichwediscussedinChapter4.2.Inacircularnetworkchart,thenetworkchartisextendedbyshowingthepositionandtheimportanceoftheobjects.Oneconcernisthatthesenetworkchartsbecomelessclearwhenmanyobjectsareinvolved,andmayevenbecomeunreadable.

Comparingdatapoints

Thebasicideaofcomparisonofdatapointsisthatoneaimstocomparescoresfor(asetof)variablesacrossmultiplesubunits(e.g.groups,time).Forexample,onemightaimtoshowthe sales per category per quarter. Or one may aim to show the conversion rates fordifferentwebsitesovertime.Again,differenttypesof“comparisoncharts”canbechosen(seeFigure4.3.9).

Figure4.3.9Comparisoncharts

If one has simple comparisons (e.g. sales per brand), then usually one of two rathersimilarcharttypesisused:

AcolumnchartisusedwhenthereisalimitednumberofsubunitsA bar chart is used when the number of subunits is larger, as more space isavailableinthisgraph.

Formorecomplexcomparisons,inwhichmultiplemeasurementsformultiplegroupsneedtobecompared,morecomplexgraphsarerequired:

A radar chart (also known as web chart, spider chart, star chart) is a graphicalmethodofdisplayingmultivariatedata intheformofatwo-dimensionalchartofthree ormore quantitative variables represented on axes starting from the samepoint(e.g.theallocatedbudgetversusactualspendingofdifferentdepartments,orscoringonproductattributesofdifferentdesigns).Sometimesitishardtovisuallycompare lengths of different spokes, because radial distances are hard to judge,though concentric circles help as grid lines. Instead, one may use a simple line

graph,particularlyfortimeseries.AbulletchartbuildsuponthebarchartandhasbeendevelopedbyStephenFew(2006:120–206).Thebulletgraphprovidesaprimarymeasureunit(e.g.year-to-yearrevenues),andcomparesthismeasurewithothermeasurementunits(e.g.target).Italso shows the context in terms of ranges of performance—for example, “good,”“average,”and“weak”(seeFigure4.3.10).

Whencomparingmeasurementsovertime,itiseasyenoughtouseacolumngraphifonlya limited number of categories for only a few periods (e.g. four quarters) is beingconsidered.However,generallymorecategoriesandmoretimeperiodsareconsidered.Aline chart is especially useful if multiple time periods are being considered (e.g. salesdevelopmentovertheyearby,say,weeklyunits).

Composition

Figure4.3.10Exampleofabulletchart

Whenusingacompositionchart,theaimistoshowhowthedataarebeingbuiltupoutofdifferentsubunits.Initsmostbasicformthisresultsfromafrequencytable.Forexample,one might like to show the distribution of the origin of website visitors over differenttouchpoints (e.g. Google, Banners, Affiliates, Direct Load). Again there are many“compositioncharts”thatcanbeused(seeFigure4.3.11).

Figure4.3.11Compositioncharts

Wewouldmakethefollowingobservationsaboutcompositioncharts:

The pie chart is very popular. It is very useful when a limited set of items orcategoriesisbeingshown.Acommonmistakeistouseitformanyitems.Thepiechartquicklybecomesunreadableandnotinformative!Someexperts(e.g.Stephen

Few) believe that one should never use a pie chart, as especially with multiplevariablestheyrequirealotofspace.Moreover,thepiechartsmightbedifficulttointerpretwithoutprovidingexactfigures.A stacked chart is a stapled columnor bar chart. It can, for example, beused toshowthedistributionofsalesperproductperregion,wheretheregionsareshownineachbarandeachbarrepresentsaproduct.Awaterfallchartisgoodforshowingthebreakdownofavariableincomponents.Incontrastwiththepiechart,thischartprovidesagoodvisualizationofthesizeofeach of the components. Moreover, it is also possible to show negative values,which is frequently impossible in many other graphs. We recommend usingdifferentcoloursforpositive(e.g.green)andnegativevalues(e.g.red).Anexampleof suchachart isabreakdownofchurneffects (i.e. total churndecomposed intoinevitable churn, price churn, bad service churn, etc.). Extensions can also showdevelopmentsovertime.Wefrequentlyalsousethischarttoshowtheexplanatorypowerofeachvariableinaregressionequation.Tag/word clouds have become increasingly popular,with the increasing usage ofunstructureddata.Usingoutcomesoftextminingexercises,theimportanceofeachwordcanbevisualizedinawordcloud.Thelargerandthemoreboldthewordthemore frequently the word is observed. Generating these clouds is nowstraightforwardusingfreeonlinetoolslikeTagul,Wordle,orTagcloud.

Someof theabovecharts canalsobeused to showchangesover time.Wehavealreadymentionedthewaterfallgraph,butthiscanalsobedonewithothergraphs:

Inthestackedcolumnchart,stapledcolumnsperperiodcanbe linkedwithsmalllinestographicallyshowthetimeelement.Thestackedareachartbecomesincreasinglypopular,asthisvisualizationcanshowmorechangesovertime.Incomparisonwithstackedcolumnchartmanymoretimeunits(evencontinuous,e.g.salesperweek/dayoverayear)canbeshown,therebyeven showing the development of multiple stacked subunits (e.g. brands in acategory). This plot can be unclear, especiallywhen there are a number of peakperiods and down periods in the data. The ratios in the data can then becomeunclear.

Distribution

Asthenamesuggests,adistributionchart(seeFigure4.3.12)isusedtodisplayhowdataisdistributedandtounderstandoutliersandcategoriesthatareoutsidethenorm.Onecould,forexample,considerthedistributionofagegroupsofcustomers,distributionofrevenuesoverallcustomers(istherean80/20rule?),orexaminethepowerofaresponseprediction

model.

Figure4.3.12Distributioncharts

Thegraphsusedtoshowdistributionaresimilartothoseusedforcomparingvariables.Graphsforasinglevariableare:

Acolumnhistogram,whichisarathersimplegraphandisusablewhenthereareafewcategoriespervariable(e.g.agegroups).A line histogram,which is similar to a column histogram but can handlemanymorecategories(e.g.agesinsteadofagegroups). Insomecasesplotscanbeused,butingeneralthesearenotrecommendedwhentherearelargepeaksanddipsinthedata(asdiscussedwiththestackedareacharts),becausethentheybecomehardtoread.

Ifaresearcheraimstodisplaymultiplevariablesandaimstoshowsomedistribution,thefollowinggraphscanbeused:

A double bar chart can be used if you want to compare, for example, thedistributionofcustomersanddistributionsofrevenues.Thisisalsocalledadecileanalysis.Liftandgainchartsareausefulwaytovisualizehowgoodapredictivemodelis.Anexamplemightbepredictingdirectmail response,whereon thehorizontalx-axisthenumberofcustomersisplotted,andontheverticaly-axis,thecumulativeliftofthepredictionmodel(seeChapter4.1).

Figure4.3.13Chartsuggestions–athoughtstarter

Source:AdaptedfromAbela(2008)

The scatter chart provides a kind of cloud of points that have a position on thehorizontal x-axis and the vertical y-axis. This may show something about thepotential association between variables, as we already noted when discussingrelationcharts.

AndrewAbela(2008)hasprovidedanicechoiceprocessforthetypeofgraphtobeused(seeFigure4.3.13).Althoughnotallthegraphsthatwediscussedhereareshown,thisflowdiagramcanbeveryusefulwhensearchingfortherightgraphtouse(seeFigure4.3.14).

Finally, we stress that our overview of graphs is not exhaustive. We have aimed todiscussthemostimportantandmostfrequentlyusedgraphs.Althoughcarehasbetakenwhenusingthesegraphs,manyofthemcancreatemoreimpactwhenincorporatedintoareportorpresentation.

GraphDesign3

After choosing the graph type one should consider the design of the graph. This is alsoessential, as the design will determine the attention the graph will get. Colin Ware,DirectoroftheDataVisualizationResearchLabattheUniversityofNewHampshire,termsthe basic building blocks of the visualization process as “pre-attentive attributes” (Ware,2008).Theseattributes immediatelycatchoureyewhenwe lookatavisualization.Theycanbeperceived in less than10milliseconds,evenbeforewemakeaconsciouseffort tonoticethem.Alistofpre-attentiveattributesisgiveninFigure4.3.14.

Figure4.3.14Pre-attentiveattributes

Source:AdaptedfromWare(2008)

These pre-attentive attributes can be useful when designing a graph, as theyimmediately identifyavisual.Theseattributesarealso thebasis forpatternsshown inagraph.

AnalyticalPatterns

ColinWare(2008)statesthat:

Ifpre-attentiveattributesarethealphabetsofvisuallanguage,analyticalpatternsarethewordsweformusingthem.Weimmediatelyidentifythepre-attentiveattributesinavisualization.Wethencombinethepre-attentiveattributestoseekoutanalyticalpatternsinthevisual.

ThebasicanalyticalpatternsaredisplayedinFigure4.3.15.

Thebasicattributesandpatternsallowreceiverstoprocessvisualinformation.However,it is not only the patterns that are useful. Beyond that onemight want to highlight oremphasizespecificpatternsoverothers.

Figure4.3.15Basicanalyticalpatterns

Source:AdaptedfromWare(2008)

Tipsandtricks

Wehave some tips and tricks tohelp the researcherworkingwithgraphs todo amuchbetterjobofcreatingthem:

Keepitsimple!Thisisthegoldenrule.Alwayschoosethesimplestwaytoconveyyourinformation.Killthegridlinesunlesstheyareabsolutelynecessary,oratleastmakethemsubtlesotheydonotdistractfromtheinformationyou’retryingtopresent.Makesureyourchartiscenteredonthedatayouwanttopresent.Your axes should be clearly labeled, and should have units on them wherenecessary,sonoonehastoguessorinferwhatyou’retryingtosay.Usecolour, size, andposition tohelp the reader to seewhat is important.Colourservestohighlightexceptions,nottoenlivenadulldashboard.Remember,yourgoalisthatanyonecanpickupyourchart,whetheryou’rethere

to talk about it or not, and understandwhat information the data are trying tocommunicate.Onefrequentlybelievesthatoutcomesofregressionmodelsshouldbeputintables.However,regressionmodelscanalsobeshowningraphs,forexamplebyshowingthestandardizedcoefficientsforthemostimportantvariablesinabarchartorbyshowing the explanatory power of each variable in awaterfall chart (see Figure4.3.11forexamples).Werecommendresearchersshouldtryoutanumberofgraphsandlearnto“play”withthem.Thiswayoneimmediatelylearnstheprosandconsofeachgraph,anditbecomeseasiertopicktherightone.Beawareofmisleadingwithgraphs.Ifdifferentscalesareusedon,say,they-axis,small effects can become visually large. Gaining attention is not the same asmisinterpretationofeffects.3D graphs can be unnecessarily confusing. The perspective information in thebackgroundcangivetheimpressionthatitislessimportantthaninformationinthefront.Usecompellingheadlinesanddeckstodescribethetake-awaymessageofthevisualization.

Trends

Weconcludewithadiscussionofsometrendsinvisualization.Inpracticeandinlinewiththe growth of big data development we observe a stronger focus on design. We alsoobserve thatmore infographics arebeingused.DavidMcCandless is taking infographicsand data-visualization to the next level. In his book Information is Beautiful (2010) hevisualizes captivating and intriguingpatterns and connections across art, science,health,andpop.Analysts frequently lack the skill for thiskindofworkandprofessionaldesignartistsarebeingused.Wealsoseean increasinguseof text-basedgraphs, suchasword-clouds.

Figure4.3.16Fromstorylinetovisualstopresentation

Weareprobablyawareofonlysomeofthevisualizationtrends.Withthetechnologicaladvancesbecomingavailabletohelpthedisplayofvisualeffects,three-dimensionalgraphswill become more popular and insightful. Just watch the way Professor Hans Roslinghandles a presentation, commentating on amoving hologram that illustrates the health,wealth, and population of 200 countries over 200 years in less than a minute (Rosling,2007).Wealsobelievethatthegrowingimportanceofvideomeansthatpresentationswillalsobecomemorevideo-based.

Conclusions

Inthischapterwehaveputforwardthepropositionthatlowimpactisageneralproblemformanyanalyticalstudies.Averyclearstorylineandvisualizationarekey-ingredientsforcreatingmoreimpact.Insum,wehavethefollowingclearrecommendationsforanalysts;if they are followed, the result should be a storyboard in which the storyline and thevisualsareintegrated(seeFigure4.3.16).

Startthepresentationbycreatingaclearstoryline.Writethestorylineinfullsentences:

–Whatisthesituation/complication?–Whatisthecoremessage?–Howcanthismessagebeunderpinned?

Continuebydrawingsomeinitialslidesandvisuals.Donotuseacomputer,butuseadrawingboardtostimulatecreativity.Writeouttheheadingsofeachslideinfullsentences.Choosetherightgraphstovisualizethesupportinginsights.Usingthisbasis,makeareport/presentation.Useacriticalcounterparttochallengethepresentation.

Notes

1Comparedtootherchapters,thischapterhasaverystrong“howto”focus.Attentiontothecreationofreportsisvery

limited in the academicmarketing literature.We have aimed towrite a comprehensive chapter that provides the

analystwithsufficientpracticalguidelinestosetupaneffectivereportorpresentation.Thisisheavilybasedonour

owncombinedexperienceofgivinghundredsofpresentationsonouranalyticalstudiesformajorcompanies.

2Thereareothermethodsaswell.However,wefindthatthismethodprovidesseveraladvantagesandfocusonitinthis

chapter.

3ThissectiondrawsheavilyonColinWare’sstudy(2008).

References

Abela,A. (2008).Advanced presentations by design:Creating communication that drivesaction.SanFrancisco,CA:Pfeiffer.

Anscombe,F.J.(1973).GraphsinStatisticalAnalysis.TheAmericanStatistician,27(1),17–21.

Few,S.(2006).Informationdashboarddesign.CA:O’Reilly.

McCandless,D.(2010).InformationisBeautiful.UK:HarperCollins.

Miller,G.A. (1956). Themagical number seven, plus orminus two: Some limits on ourcapacityforprocessinginformation.PsychologicalReview,63,81–97.

Minto, B. (2009).The Pyramid Principle: Logic inwriting and thinking. EdinburghGate,Harlow,Essex:PearsonEducation.

Paivio,A.,&Csapo,K. (1973).Picturesuperiority in free recall: Imageryordualcoding?CognitivePsychology,5,176–206.

Roberts,J.H.,Kayandé,U.,&Stremersch,S.(2014).Fromacademicresearchtomarketingpractice: Exploring the marketing science value chain. International Journal ofResearchinMarketing,3(2),127–140.

Rosling,H.(2007).Newinsightsonproperty.RetrievedfromTED.com.RetrievedSeptember11,2015fromwww.ted.com/talks/hans_rosling_reveals_new_insights_on_poverty?language=en

Ware, C. (2008). Visual thinking: for design. Morgan Kaufmann Series in InteractiveTechnologies.Amsterdam:Elsevier.

Wurman,R.S.(1989).Informationanxiety.NewYork:Doubleday.

5Buildingsuccessfulbigdatacapabilities

Introduction

Havingbigdata is not sufficient of itself to develop a successful value-creatingbigdatastrategy.Firmsneedtoinvestinbigdatacapabilitiesthattransformtheorganizationintoafact-basedandmoredata-drivenenterprise.Atfirstblushthisseemseasytoachieve.Firmsshouldjustbuysomesoftware,hiresomebigdataexperts,andthebigdatainitiativescantake-off.AccordingtoIDG(2014)thetopfivebigdatainvestmentsinvolvestorage,servers,cloud infrastructure, discovery and analytics, and applications. However, many of theseinvestmentswillcertainlyleadtoseriousdisappointments.AndindeedaccordingtoCisco60%ofcompaniesagreethatbigdatawillhelpimprovedecisionmakingandincreasetheircompetitiveness,1butonly28%indicate that theyarecurrentlygeneratingstrategicvaluefromtheirdata.Probablysomeshort-termsuccessescanbeachieved;however,foralong-term impact firms should invest inpeople, systems,processes,and theorganization.Ourexperienceshaveshownthatfirmsmayfaceseveralhurdlestodoso.Forexample, firmsmay be confronted with old systems and databases that are difficult to replace.Importantly, there is also a shortage of analytical talent (Manyika et al. 2011; Leeflang,Verhoef,Dahlström,&Freundt,2014;seealsoFigure5.1).Findinganalyticalpeopletohirecanbe a nightmare.WesternEuropean companies arenowhiring talent from India andotherAsiancountriestofillanalyticalvacancies.Anotherhurdleistochangetheculturetoone in which analytical solutions are considered as very valuable input in marketingdecisionmaking,insteadofonlyfocusingonintuition.

Inthischapterwewilldiscussthemajoringredientsfromwhichfirmscansuccessfullybuild an analytical competence, one that fully allows them to benefit from big dataopportunities.Basedonourbigdatavaluecreationmodel,wewillstructureourdiscussionmainlyaroundfourbuildingblocksofasuccessfulanalyticalcompetence:

1. Process:creatingacommonbusiness-drivenwayofworkingthatisunderpinnedbyprivacyandsecurity

2. People:torecruit,todevelop,andtomaintaintherightanalyticaltalent3. Systems:theplatformandtoolsforanintegrateddata-ecosystem4. Organization: taking the right role and place in the organization for the most

impact.

However, we will start with our vision on what the transformation of an organizationentailsifitwishestocreatestronganalyticalcompetences.

Figure5.1Shortageofsupplyinanalyticaltalent

Source:AdaptedfromManyikaetal.(2011)

Transformationtocreatesuccessfulanalyticalcompetence

Over the past few decades marketing intelligence (MI) has become a more importantfunction within many firms. We observe this especially in service industries, such asretailing, telecom, and financial services. Numerous examples have been discussed inliteratureandcanbeseeninpractice, inwhichorganizationssuchasTesco,CapitalOne,O’Hara’s,andKPN,havedevelopedanalyticalfunctionstohelpthemeffectivelycompete(e.g.Verhoef&Lemon, 2013;Davenport&Harris, 2007;Humby,Hunt&Phillips, 2008).And as already discussed there is also ample scientific empirical evidence that firmsinvesting in these analytical competences can actually outperform competition (e.g.Germann, Lilien, & Rangaswamy, 2012; Germann, Lilien, Fiedler, & Kraus, 2014). MIdepartmentsorfunctionsperformanimportantrole increatingthesecompetences. Inaneraofbigdatathiswillbecomeevenmoreimportant.

However,wealsoobservethatinmanycompaniesanalyticalcompetenceshavenotbeenfullydeveloped.Infactonecouldputforwardthestatementthatmanyfirmsarenotwell-prepared internally for the presumed big data revolution. They simply lack a stronganalyticalcompetence.Asaconsequence,theMIdepartment(ifpresentatall),isnotlikelyto play an important driving role in big data value creation. In many firms thesedepartments stick to their traditional role, inwhich they frequently only deliver simplereports and customer selections on requests and do not actively participate in value-creation discussion lacking strongmarket and customer insight—all because of a lack ofanalyticalcapabilities.TheMIdepartmentthenmainlyhasareactivesupplierrole.

Changingroles

IngeneralwehaveobservedachangingroleinhowMIoperatesinfirms.Thesechangingroles are displayed in Figure 5.2. MI departments frequently start in a supplier role inwhichthecredois:“Wedeliverwhatisbeingaskedfor.”Inanextphasetheycanbecomeachallenger, inwhich theyarealsoasked toprovide inputonspecificmarketingdecisionsbasedontheirpresumedmarketandcustomerknowledge.Inthenextphases,thefunctionmoves froman advisor role, to an initiator role, and finally anorchestrating role.Thesethreephasescanbeconsideredas top-classMIroles. Intheadvisorrole, theMIfunctiongetsastrongersayinmarketingdecisionmakingandconsultsitasanimportantadvisor,whose input is valued and strongly taken into account.As an initiator, theMI functiondevelops independent marketing proposals in close cooperation with the marketingdepartment.Theorchestratingroleisobservedveryinfrequently,andbasicallyimpliesthatMI is embedded within marketing decision making and becomes a driving forceresponsiblefororchestratingcustomercontactsoverthemultipleavailablechannels.

Figure5.2Changingroleofthemarketingintelligencedepartment

Source:AdaptedfromVerhoef,Hoekstra,&VanderScheer(2009)

Changingfocus

The changing role of theMI department induces a changing focus, scope, and availablecapabilitieswithinMI department. Based on our years of experiencewith organizationsthathavemadeastrategicchoice tobring the intelligence function toahigher level,westate anoutline of thedesired changes and resulting outcomes these organizations oftenenvision and experience (seeTable5.1). In sum,we observe that functionsmove from atacticalfocustoamorestrategicfocus.Asaconsequence,theirinputindecisionmakingischangingandtheyalsouseadifferentanalyticalapproach.Insteadoflooking-back,amoreforward-looking predictive approach is embraced. They also consider that an integratedviewon customer andmarket data is required. In doing so, theywill be a driving forcewithin the organization to stimulate data integration developments. In terms of dailyoperations, we observe a shift from reactive analyses to more proactive agenda-settinganalyses.Thisalsoinducesastrongfocusonclearreportsandvisualizationoftheresults.Overall,thesechangeswillresultinanattractiveplacetowork,whereyoung,talented,andambitiousanalystsarewillingtowork.

ForasuccessfulexamplewediscussthetransformationoftheMIdepartmentatDutchTelcoKPN(seeBox5.1).Amainchallengeforfirmstoachievethesetransformationsisindevelopinganintelligenceteamwiththerightskills.Wewillelaborateonthatinthenextsections.

Table5.1Shiftingfocusofthemarketingintelligencefunction

Area From To

Strategicfocus Tacticalandshort-termfocusonactionsandcampaigns

Strategicandlong-termfocusonV2CandV2F

Inputmarketingdecisionmaking

ProvideroffiguresandqueriesFact-basedandactionableadvicethatmeetsthebusinessplanning

Analyticalapproach

Lookingbackandexplanatoryinsights

Forward-lookinginsightsandconcreteproposalsforchange

Intelligencerole

Fragmentedviewofthecustomerbaseonscattereddata,information,andknowledgeacrossmultipledepartments/silos

One-intelligenceapproachandmethodwithasinglecustomerviewthatisaccessiblethroughouttheorganization

Dailyoperations

Eliminatingworkload,whichisfilledbythedailyoperation,(reactive)isthenorm

ProactivelysettingownagendaandprioritiesinlinewiththeorganizationalKPIs

Output SupplyofrawoutputinExceltypeofprogram

Clearandstrongvisualizedpresentationswithclearmessage

Visibilitywithinfirm

Impactvisiblefortheimmediatearea(i.e.marketingandsales)

Impactvisiblefortheentireorganization

Attractivenessoffunction

Departmentwithlimitedgrowthprospects

Departmentinfluxofnewtalentwithattractivecareerpathsandabreedingsitefortalent

Attractivenessforemployees

Obscurityasanattractiveemployerforitsleadinganalytics

Preferredemployerforanalyticaltoptalent

Box5.1BuildingupananalyticalcompetenceatKPNThe Dutch incumbent Telco KPN understood the necessity of building up anintelligence function several years ago. Confronted with suboptimal marketingcampaigns, increasing competition, and high increasing churn, the firm beganimplementing customermanagementwith a strong focus on fact-basedmarketing.Themanagement ofKPN also understood that a strongMI functionwas essential.However, the firm lackedstronganalytical skills.Therefore, it setup theMarketingIntelligence(MI)AcademyjointlywiththemarketingintelligenceconsultingagencyMIcompany.BaptiestCoopmans,formermemberofKPN’sboardofdirectors,stated:“Ourowneducationprogramswerenotfocusedoncustomerintelligence,butfocusedoneducatingmarketingmanagersandorganizational leaders.With theMIacademywe aim to develop top talented specialists, which are able to fully understandcustomerneeds.”Athree-yeartrainingprogramwasdevelopedfornewyoungmastergraduates.Inthisprogram,theyweretrainedinanalyticaltechniquesbutwerealsotrainedinimprovingtheinterfacebetweentheMIfunctionandmarketing.Importantskills included the effective presentation of marketing facts and interpersonalcommunication. Furthermore, MI trainees were trained in deriving growthopportunities from their analytical findings/insights to develop strongrecommendationsformarketing.Thewholeprogramhasbeenevaluated(seefigure

below). The management of KPN is fully convinced of the success of building astrong MI with MIacademy. Marteyn Roose, at that time director of customermanagement at KPN, stated: “Statistical tooling, methods and central datawarehousingareallbeneficial,butit’sourpeoplethatmaketherealdifference.Thebenefits of attracting top talented graduates and training them to become topperformingMIprofessionalshaveproventobetremendousinallkindsofareas.”KPNidentified several of the benefits of the program shown below. Since then, theprogram has resulted in additional revenues and cost reductions. Direct revenueshaveoccurredthroughimprovedmarketingcampaigns(overallimprovingnetpresentvalue of orders by 132%) and the development of successful growth initiativeopportunities(>€140millionin2010).Indirectandmorelong-termrevenueshavealsobeen realized, for example, through the strong development of the customermanagementfunctionwithinKPN.ThereisanincreasinglyhighdemandthroughoutalldepartmentsanddisciplinesinKPN.ThisensuresthatfewMIprofessionalschooseto leaveKPNto lookforchallengingcareerselsewhere.Thetraditional focusofMIwas mainly the optimization of marketing campaigns. At KPN, that scope wasbroadenedtoalldisciplines(e.g.,marketing,sales,service,network)andonalllevels(i.e.,uptotheboardofdirectors).AstrongMIfunctionisanecessity.Theprogram’sstrong focus on the interface between marketing and MI also resulted in greateracceptance of more accountable or fact-based marketing. Cost reductions mainlyoccurredfromlowerrecruitmentcostsandshortvacancyperiods(>95%ofnewtalentacquiredwithoutanyrecruitmentaid).Moreover, the firmcouldpay lowersalariesbecause it did not need to hire experienced MI specialists (20% saving, includingeducational costs). The firm also required less input from external dedicatedmarketingintelligencefirms,assufficientskillswerenowpresentinternally.

Source:Verhoef&Lemon(2013)2

Buildingblock1:Process

The “process” focuses mainly on how, within firms, analytical projects are defined andexecuted.Weconsiderfiveimportantphasesoftheanalysis(seeFigure5.3):

Figure5.3Phasesofthestandardanalyticalprocess

Thisprocessisdrivenbythebusinessquestionandleadstooutputwhosepurposeistoanswerthatquestion.Asoutlined,thisprocessseemsratherstraightforwardandfitswithour preferred analytical problem-solving analysis strategy.However, themost importanthurdleisthecooperationofandunderstandingbetweenthedifferentdepartmentsandtheMIdepartment. These departments can sometimes function as separate silos that donotunderstandeachother(seeBox5.2).ThismaycomewiththeriskthatanalyseslosetheirusefulnessandtheMIdepartmentbecomesobsolete.Itcanbecomeevenmorecomplicatedwhenthereareseparatedepartmentsgainingmarketandcustomerinsights.

Box5.2ViewsonthemarketingintelligenceprocessMarketingIntelligence:

“Marketingdoesnotunderstandthatansweringasimplequestiontakesalotofworkforus”“Werecentlydidagreatanalysis,butnothingisdonewithit”

Marketing:

“Marketing Intelligence has written a wonderful extensive report, but thereportdoesnotreallyanswermyquestion”“Iamconvincedthattheresultsaresignificantlyreliable,butIdonotseehowIcanapplythisinpractice”

Sales:

“Marketing Intelligence supports the Marketing department with itscampaigns,buttheytotallydonotunderstandwhatisgoingoninourfield”“Inaddition,theydonothaveaccesstotherelevantdataforus”

Finance:

“As long as the figures of Marketing Intelligence are not in line with our

reporting, findings are unreliable and therefore not useful for control andpolicy”

The opinions given in Box 5.2 reflect how the MI department functions within theorganization, and how thismay result in reduced impact. There are threemain reasonswhytheintendedimpactandaddedvalueofanalyticsisfrequentlylimited:

UnclearstartingpointoftheanalysisLimitedconnectiontothedynamicsofthebusinessplanningprocessUnclearandnotimpactfulpresentationofanalyses.

Startingpointoftheanalysis

Fordecades,analystsweremainlyusedtoprovideinformationthatfocusedonprovidingrather descriptive analyses giving explanations, making comparisons, and evaluatingmarketingcampaigns.Typicalquestionsformarketinganalystsare:

Howmanyactivecustomersdowehave?Whatistheprofileofnewcustomers?Whyhastheoutflowofcustomersincreasedoverthepastperiod?Whatsegmentscanwedistinguishinourcustomerdatabase?Whatwastheresponsetothelastcampaign?

There is nothing wrong with asking these questions. However, it is not clear what theunderlying business challenge actually is. Focusing solely on the above questionswouldhavesomeimpactonmarketingdecisions,butwouldnothaveanimpactthatgoesbeyondmarketing.Inordertobemoreimpactfulthestartingpointoftheanalysisshouldalwaysberelatedtoaclearbusinesschallenge.Thisgivestherightfocus,therightpriorityforuseofscarceresources(time,capacity,andbudgets),andcreatesacommonacknowledgementof the relevance and importanceof the executed analyses.Weobservedone firmwithinwhichforyearsthemarketingresearchdepartmentexecutedanalysesonhowtoimprovesatisfaction—withouthavingmuchimpact.Atsomepoint,satisfactionbecameakey-metricrequiringattentionforthefirmanditsstakeholders.Theintelligencedepartmentstartedtounderstand the relevance of the metric to the business challenge and started impactfulprojects, supported by top management, on reporting, explaining, and predictingsatisfaction. It is thus important for analysts tounderstand thebusiness challenge.Goodbusinesschallengesshould:

AlwaysbelinkedtoobjectivesandKPIsandthereforemeasurableShould be linked to themes as market attractiveness, brand performance, andcustomermanagement(seealsoChapter2)

Resultindifferentactionableoptionsformanagement.

Hereareanumberofexamplesofclearbusinessquestionswhichgivetherightfocusfortheanalysisprocess:

Howcanthechurnofpost-paidsubscribersintheconsumermarketbereducedby3%inthecomingyear?How canwe increase the cross-sell ratio of our insurance products for the retailmarketfrom1.2to1.8?How can we increase our market share by 5% through the acquisition of newcustomersinthebusinessforsmallenterprises?How dowe have to allocate ourmedia budget to increase our brand preferencefrom10%to15%?

Togetawell-definedandto-the-pointbusinessquestiondrivingtheanalyses,webelievethefollowingissuesarecrucial:

Haveagoodinitialdiscussionwithrelevantownersoftheproblemwithinthefirm.Clearlydetermineonwhich levelyouranalysishas the strongest impact:market,brand,orcustomer.Determinewhethertheprojectfocusesonincrementalimprovementoroptimizingof current business andmarketing practices or aims to achieve strategic changesandgrowthopportunities;DeterminewhichdefinitionsoftheanalyzedKPIsareusedwithinthefirm.Assess the importance of the identified business challenge by quantifying thebusiness impact of the problem (e.g. reducing churn could lead to an increase inEBITDA).DeterminetheintendedchangeintheKPIofinterest(i.e.reducechurnwith3%).Validateandsettargets/objectivesbasedonexistingbusinessplans,strategypapers,outlooks,andreports.Makesure thatyouhavedifferentpropositionsandorsolutions for the identifiedbusiness challenge thatneeds to testedor evaluated in suchawayas tokeepanopenmindtoperhapsunexpectedopportunities,ratherthanimmediatelycomeupwithakindofstandardandfrequentlyexpectedadvice.

Supportduringtheanalysisprocess

Oncethebusinesschallengeisdefined,thenextsteptoensureimpactduringtheanalysisisgoodmanagement.Wehavefoundthefollowingtobeusefulguidelines:

Ensureconsensusonthebusinessdemand(withbusinessmanagersandfinance&

control).Makesuretherightpeopleareinvolvedandinformedduringtheanalysisprocess.Provideascheduleandoverviewofthestepsthatyougothrough.Ask foractive supportduring theanalysisphase, including timeandcapacity forcoordination,consultation,andfeedback.Makesurethatthereissufficientaccesstoallnecessaryinformation.Make sure that there is budget for (unexpected) out-of-pocket expenses (insertexternaldata,capacity,etc.).

GivenwheretheMIfunctioncomesfrom,itisnotsostrangethatmanyintelligenceteamsarenotfullyawareofwhatthemostimportanttopicsfortopmanagementare.Tohelpgetaprojectrollingandtogetmoresupportfromtopmanagement,knowingtheiragendaisofgreat importance. Itwillalsohelp toask the rightquestions.Sowestronglyrecommendthat MI specialists look beyond daily tasks and marketing challenges. A relatedrecommendationisthatanalystsshouldbeconnectedwithfinance.InmanyorganizationsfinancehaveastrongsayinsettingKPIsandthedefinitionoftheseKPIs,astheyhavetobereportedtoexternalstakeholders.Hence,whenanalyzingthesemetricsanalystsshouldusethemetricsasbeingusedbyfinance.Analysesinwhichmetricsdifferfromthoseusedin finance areoften interpretedasnot reliable and thereforenotuseful.Alwaysvalidateresults based on existing outlooks, business plans, strategy documents, andmanagementreports.

Buildingblock2:People

The transformation of theMI department also requires the attraction of employeeswhocanworkeffectivelyinthesedepartments.Thelargestchallengeisthatitisdifficulttofindgoodemployees.Thereisashortageofgoodanalyststhatcanfunctionwellinthishighlydemandingjob.Includingthismarketchallengetherearethreepeoplechallenges:

1. Selection:Whatistheprofileoftheemployeesrequired?2. Acquisition:HowdoIacquiretherightanalystsfromthejobmarket?3. Retention:HowdoIkeepthegoodanalysts,giventheshortageinthemarket?

Analystprofile

Althoughwearemainlyemphasizingthefunctionanalyst,thisseemslikeanold-fashionedterm.Intheeraofbigdata,firmsarenolongerlookingforanalysts,butfordatascientists.Thedata scientist is then consideredasbeing the sexiest job in theworld (Davenport&Patil,2012).Thereareseveraldescriptionsofdatascientists,forexample:

Ahighrankingprofessionalwithtrainingandcuriositytomakediscoveriesintheworldofbigdata

(Davenport&Patil,2012:71)

Adatascientistissomebodywhoisinquisitive,whocanstareatdataandspottrends.It’salmostlikeaRenaissanceindividualwhoreallywantstolearnandbringchangetoanorganization

(BhambhriIBM)3

Inourviewthedescriptionsarerathervagueanddonotsufficientlyfocusonwhatfirmsshould actually look for and cannot be used to select and/or train people.We thereforefocus on the skills an MI analyst or a big data scientist should have. These skills orindividualcapabilitiescanbedividedintofourareas(seeFigure5.4):

AnalyticalcapabilitiesDataandtoolsBusinesssenseCommunicationandvisualization.

AlltheskillslistedinFigure5.4areimportant.However,unfortunatelytheseskillsarenotoftencombinedinasingleperson.Forexampleaspecialistwithexcellentanalyticalskillsfrequentlylacksbusinesssenseandfindsitverydifficulttocommunicateeffectively.Thisdoes not imply that one should not hire top analytical talent! These experts are needed,especiallygiventoday’sstrongbigdatachallengesandthesophisticatedanalyticalmodelsrequiredtosolvethem.Anotherstrategywouldbetogoforsomeonescoringaverageonevery dimension. However, it is unlikely that these people will move the organizationforward,astheremightbealackofinnovationoneverydimension.Intermsofbuilding

ananalyticalcompetencewiththerightpeople,firmsshouldstrivetodevelopagoodteamofprofessionalsinwhicheachofthesecapabilitiesissufficientlyprovidedatahighlevel.Wehavesomespecificrecommendations:

Figure5.4Multi-disciplinaryskillsofananalyst

Donotaimtoget theemployeethathasall thesecapabilities (youcouldcall thisthe“sheepwithfivelegs”syndrome).Insteadbuildupateaminwhicheachofthecapabilitiesispresentatasufficientlyhighqualitylevel.Do not aim to build a fully fledged analytical capability all at once but take astepwiseapproach.

Teamapproach

Figure5.5Possiblebigdatastaffprofiles

Tohelpsetuptheteamapproach,itcanhelptomapMIemployeesontheirspecificprofile.

Basedonthefourcapabilities,onarathergenerallevelfourprofilescanbedistinguished(seeFigure5.5):

1. Theconsultant:Strongbusinesssenseandcommunication&visualizationskills2. Thedataspecialist:Stronganalyticalanddata&toolsskills3. Thedataanalyst:Stronganalyticalandspecificallyvisualizationskills4. TheITprofessional:Strongdata&toolsandbusinesssenseskills.

Importantly, buildingup the analytical team shouldusually bedone rather gradually, asshowninFigure5.6.Frequently,firmsstartbyusingtheservicesofsomeindependentdataanalysts and/or specialized consulting agencies. At some point the need for their ownanalyticalteambecomesapparentandasmallteamisbuiltup.Thisteamshouldthenbeeducatedusingspecificallydevelopedinternalprogramsorexternalprogramsofbusinessschools and/or consulting agencies. At some point there should be sufficient knowledgeand expertisewithin the firm, and firms should be able to independently build up theirownanalyticalcompetences.Atthisstageitbecomesimportanttodevelopcareerpathsfortheanalystsintheteam.

Figure5.6Stepwisedevelopmentofanalyticalcompetencewithinthefirm

Acquiringgoodpeople

Thesupplyofgoodanalystsismuchlowerthanthedemand(seeFigure5.1).Thiscreatesahugechallengeforfirmsaimingtobuildananalyticalcompetence.Onecaneasilydeveloptheidealdepartmentintermsofrequiredprofiles.However,gettingtherightpeopleisfarfrom trivial. One of the problems is that MBA programs are not good at developinganalyticalskillsastheymainlyfocusonmanagerialskills.WecanfindfewMBAprogramsthat seem to focus on big data. Also regular Master of Science programs in marketingtypicallyfocusonmarketingmanagementknowledgeandprovidelimitedopportunitiesforstudyinganalytics.Inachangingandmoredata-drivenmarketingenvironmentastronger

focusonanalyticcourseswithinmarketingprogramsbecomesanecessity.Thesupplyofstudentswell trainedinstatistics isalsolimitedinmanyWesterncountries,wheremanystudents prefer to study psychology, management, law etc.4We also observe that well-trainedstudentswithstrongeconometricskillsfrequentlyprefertoworkinfinanceratherthaninmarketing.Oneofthereasonsforthisisthatmarketingisfrequentlyconsideredasnot being quantitative in nature; another is that salaries in quantitative finance arefrequentlyhigher.Thereisajobtodoforfirmsaswellasmarketingdepartmentswithinuniversitiestoconvincethesestudentsthatmarketingisactuallyaveryattractiveworkingarea. An example of good practice is Erasmus University, where the econometricdepartment,ontheinstigationofProfessorPhilipHansFranses(ProfessorinEconometricsand Marketing Research), started a marketing track for econometricians more than adecadeago.Thefollowingsuggestionsarehelpfulforattractingtopanalyticaltalent:

Anobvious suggestion is that firms aim to buildup employeebrand equity (BE)(Tavassoli, Sorescu, & Chandy, 2014). Young professionals prefer companies, likeGoogle,thatareinnovativeandhaveastrongBE.Providinganactiveeducationplanforyounganalystsmayhelpaswell.Companiesprovidingtheseeducationprogramsinwhichyounganalystsaretrainedonthejobhaveauniquesellingpoint.Reachouttouniversitiesandbusinessschoolsinwhichtalentedyoungpeopleareeducated.Companiescouldparticipate inanalyticalclasseswithdata-basedcases,inwhich theymeet students, and students become familiarwith companies andtheiranalyticaljobprospects.

Beyondattractingyoung talent, firmsalsohavesomeother solutionsat theirdisposal togetaroundtheshortageofanalyticaltalent:

Onesolution is to trainexistingemployees in theareaofmarketingresearchanddatabasemarketing.Ourexperience,though,isthatthisisnotaseasyasitsounds.Knowledgeoftheseemployeeshastosomeextentbecomeobsoleteinabigdataeraand a training of new work mentality with a stronger focus on business,communicationandvisualizationisnotstraightforward.Asecondsolutionistomakesureeffectiveworkprocesseswithinorganizationsareinplaceandtoautomatespecificprocesses.Asaconsequencelesstalentisneeded.Forexample,customerselectionscouldprobablybeautomated.Alsoreportingcanbeautomated,especiallyforcontinuousdatacollection.

Talentretention

Giventheshortageoftalentandtheinvestmentsrequiredineducationofacquiredtalent,employeeretentionisofurgentimportanceforanalyticalfunctions.Furthermore,analysts

builduptacitknowledgethatisveryvaluabletokeep.Henceanalystretentionshouldbeatop priority for firms building up an analytical competence. Holtom, Mitchell, Lee andEberly (2008) warn that in industry the chronic shortages of qualified employees havedrivenupthecostsofturnovermuchfasterthantherateofinflation.Thisunderlinestheimportance of the loyalty of talented people. Frequently, loyal analysts are not the topemployeeswithinthefirm.Sothereisadangerthatiffirmsdonotinvestintheloyaltyoftalentedpeople, theirMIfunctionmayslidedownthe ladderandtheneedtobuildupafully fledged MI department starts all over again. We have seen that happen in manyorganizations. For example, within a service firm, at some point talented employeesbecameunsatisfiedand,asaconsequenceofalackofsupportfromtopmanagementandongoing uncertainty, started looking around, and found jobs in other industries, such asonline retailing and finance. For talent retention, several issues have been deemedimportant,suchassalariesandatmosphere,butgiventhefocusonyoungtalent,themostimportantfactorisprobablypersonaldevelopment(e.g.DeVos&Meganck,2009).Wehavethefollowingsuggestionsforfirms:

Develop attractive future career opportunities within MI departments, butpotentiallyalsooutside thedepartment. If formerMIemployeesbecomeactive inotherfunctions(e.g.finance,marketing),theymaybecomeactiveambassadorsforbigdataapproacheswithinthefirm.Createsufficientopportunitiesforpersonaldevelopmentandfreedomtoinnovate.Data-scientists are looking for ways to use data to solve problems, and are lessattractedbyspecifictasks,suchascustomerselections.Importantly, a corporate vision on big data usage and building up the analyticalcompetence is an important ingredient for creating a working environment inwhich talentedpeople feel they can contribute to the success of the organizationandcareeropportunitiesaresufficientlywarranted.Organizationliteraturereferstothis:thereshouldbesufficientorganizationalsupport(Holtometal.,2008).

Buildingblock3:Systems

Oneoftheinitialreactionsoffirmswithregardtobigdatacouldbeanimmediateaimtobuildlargesystemsinwhichallbigdataareintegratedandanalyticalandsupporttoolsareincludedaswell.Insuchawayonelargebigdatasystemiscreated.Thisapproachwouldbein linewithwhatwehaveseenwhenfirmswere implementingcustomerrelationshipmanagement. Firms invested millions in integrated software systems such as Oracle,Microsoft,Siebel,etc.Oneofthemajorlessonslearnedwasthatfirmsshouldnotfocusonthetechnologybutshouldfocusonwhatthetechnologycandoforcustomers(Verhoef&Lemon,2013;Rigby,2014).Thismajorlessoniseasilyforgotteninabigdataenvironmentwhere every IT-oriented consulting and tooling firm is communicating the impressiveopportunitiesofbigdata.Forfirmsthereisaplethoraofoptionstochoosefromdifferentproviders and software solutions. In the 2015 edition ofmarketing technology landscapesupergraphicofchiefmartic.com5 1,876 vendors are represented across 43 categories.Thenumber of vendors nearly doubled from the previous year’s edition, which charted 947companies.Andthenumberofvendorsisstillgrowing(seeFigure5.7).

Figure5.7Numberofvendorsinmarketingtechnologylandscaperepresentedinsupergraphicsofchiefmaric.com

As we have already discussed in earlier chapters, sound expectations should drivebusinessdecisionsonbigdata.Thisalsoholdsforsysteminvestments.Soitisalsotruethatwithbigdatasystems,technologyshouldneverleadthedecisionshere!Withregardtobigdata marketing solutions, the systems should be user and customer-driven instead oftechnology-driven. In that sense marketing analysts should have a strong say in thedevelopmentofbigdatasystemsandmarketingshouldbeanadvocateofthecustomerinthisprocess.JointlywithITtheyshouldcomeupwithworkableandscalablesolutions.

Weobservemanyfirmsstrivingforatotalnewbigdatasystemwhichisfrequentlyjustnotfeasibleandnotnecessary.Manyfirmsalreadyhavegoodsystemsinplacefordifferentsolutions(e.g.CRMorenterpriseresourceplanning).Theimportantthingaboutabigdataenvironmentisthatsystemscanbelinked.Weadvocateanapproachinwhicholdsystemsandnewbigdatasolutions(e.g.Hadoop)co-existwithinanorganization, looselydividedintofourdifferent‘layers’(seeFigure5.8):

1. Datasources2. Datastoring3. Analyticaldataplatform4. Analyticalapplications.

It is thus very important to emphasize that the use of one big system is generally anillusion.Firmswillusedifferentsystemsinthedifferentidentifiedlayers,butevenwithinthelayers(e.g.analytics),differentsoftwarecanbeuseddependingontheobjectivestobeachieved.

Figure5.8Differentlayersofabigdataanalyticalsystem

Datasources

Withinafirmtherearemultipledatasources.Manyofthesedatasourcesfunctionontheirown.Forexample,afirmcanhaveagoodbillingdatasystem,whichfunctionsprettywellfor billing purposes. Similarly, they could have databases in which operational data oncustomerservicerequestsarestored.Beforestartingabigdatasystem, firmsfirstshouldmakesurethatthedata ineachofthese internaldatabases isgoodandreliable.Bigdataanalysesareuselesswhenbaddataareanalyzed,despitethefactthattheanalyzeddatasetsarelarge.Inthisrespecttheoldwisdom“garbageinisgarbageout”holdsinabigdataeraaswell.Inmanyfirmstherehavebeeneffortstointegratedatabaseshavingasimilarmainfocus.SpecificallyfirmshaveinvestedinCRMdatabasesinwhichbillingdata,marketingdata, customer service data etc. are integrated to create a total view on each singlecustomer. Integrated customer databases are of essential importance for customer levelanalyses and are a key ingredient of successful predictions on churn, lifetimevalue, etc.Traditionally,thesedatabasesarethenavailableinadatawarehouse.Althoughthesedatawarehouses could involve millions of data points depending on firm size, number ofcustomersetc.,theyarerelativelysimplecomparedtotherequirementsofthenewbigdataeracurrentlyconfrontingfirms.

Asalreadydiscussed, big data development leads to largedata volumes, but probablymoreimportantisthatthereareverydifferentdatasourceswithdifferentdatastructuresanddatatypes,whichcannotbelinkedeasily.Forexample,whereasserviceinteractionsinacallcentercanbelinkedtoindividualcustomers,interactionsonwebsiteswithdifferentdevices aremore difficult to link to an individual customer (see Chapter 3.1). Similarly,unstructureddataonwebsites,blogs,andsocialmediacannotbefullylinkedtoindividualcustomers.Furthermore,aggregationlevelsofdatamaydiffer(e.g.brandlevelvs.customerlevel).Itisimpossibletostorethesedatainanintegrateddatasystem.Westronglyadvisethat inabigdata strategy, firms shouldnot strive to integratealldifferentdata sources.They should only integrate data that can easily be integrated because there are strongidentifiers between the data sources (e.g. customer id for CRM database). Instead theyshould create a big data platform in which the several databases are available and caneasilybeaccessedbyanalystsbasedontheirrespectiveanalyticalquestions.

Datastorage

Several data sources can be saved in a large data warehouse. This data warehouse istypically internal to the firm. We refer to Chapter 3 on the specifics of data storage.Importantly, many firms now also use the cloud to store data. The typical way ofprocessing and integrating all data sources that have been assessed as relevant is calledETL (extract/transform/load): it handles all input data sources in order to save and usetheminthedatawarehouse(seealsoChapter3.1).

Analyticalbigdataplatform

The big data platform thus involves a set of databases which should be accessible byanalyticalspecialists.Thesedatabasescanbethemoretraditionalsortavailableinstandarddata warehouses, but for large volumes and/or unstructured data, specific big datasolutionssuchasHadoopclusterscanbeused.Firmsshouldtakecareofsecurityissuesandshouldonlyprovideauthorizationtoaccessspecificdatatointernallyselectedemployees.Wemustemphasizethatthisbigdataplatformisonlyadataplatform.Usingthisbigdataplatform a database specialist could aim to combine specific data sources through, forexample,datafusionandSQLqueries.Forexample,whenstudyingthechurnofcustomers,one could combineCRMdatawithdataon socialmedia (e.g. likes for specific customersegments). Or analysts, when doing longitudinal analyses of brand sales, could build adatabaseconsistingofweeklybrandsales,weeklyadvertisingefforts,weeklysocialmedialikes,andthedigitalsentimentindex(DSI).

These datasets should be structured in such a way that they can be analyzed in theprogrambeingused. Inpractice the term“datamart” isused to refer todatasetsderived

from the big data warehouse. A data mart is the access layer of the data warehouseenvironmentthatisusedtogetdataouttotheusers.Datamartsaretypicallysmallsubsetsofthedatawarehouseandareusuallyorientedtoaspecificbusinesslineorteam,inthiscase the MI team. Whereas data warehouses have an enterprise-wide depth, theinformationindatamartspertainstoasingledepartment.Fromthisdatamartflatfilescanbeextractedwhichcanbefurtherintegratedandprocessedtobeusedinanalyticaltools.Often these analytical tools alsohave functionalities for further data processing, such ascreatingallkindofnewvariablesbyclassifyingorcombiningthedifferentinputvariables.

Analyticaltoolsorpackagesfocusonthestatistical,econometric,andlinguisticanalysisofthedata.Inthepast,firmstypicallychoseapreferredanalyticaltool.Thistoolcouldbeeither (partially) specifically developed for the firm or a standard statistical softwarepackage,suchasSPSS(nowadayspartofIBM)orSAS.Aspecificallydevelopedtoolwouldconsist of typical analyses being used in a firm presented in a user-friendly way. Datainflowtotheseprogramsisthenfrequentlyautomated.However,thisapproachhaslimitedscope and it is virtually impossible for an analyst to do additional analyses on data.Statistical packages have a full set of statistical analyses available that can be used formultiple analytical purposes. For example, in packages like SPSS and SAS standardavailabletechniquesincludealldescriptiveanalyses,suchasregressiontechniques,clusteranalysis, etc. Analysts are often trained in these packages in their bachelor andmasterprogramsatuniversities.InfactoneofthestrategiesofSPSShasbeentoprovidesoftwareto students at lowprices. The use of these packages providesmore freedom to analysts.However, a sufficient understanding of this software and the underlying techniques isrequired.Oneofthemistakeswefrequentlyobserveisthatsoftwareusersstartanalyzingdatawith(advanced)techniqueswithoutasufficientunderstandingofthedifferentpitfallsofdoingso.

Despitetheirlimitations,thediscussedsoftwarepackageshaveprovedtobeveryusefuland have becomemore user friendly over the years,with the introduction ofwindows-interfaces,helpfunctions,etc.Thebigdatadevelopmentischangingthestatisticalanalysisfield dramatically. The statistical packages are not yet fully suited for big data analytics(seealsoChapter4.2).Wespecificallyobservethefollowingdevelopments:

Due to the presence ofmore unstructured data, new analysis techniques such astextminingarebeingusedmoreoften.Frequently, statistical packages are limited in terms of the data volume they canhandle.Newadvancedtechniques,suchasBayesianandextensivepanelanalyses,arenotavailableinthesepackages.Thebigdatadevelopmenthasattractedbigdatascientists,whoarecreatingtheirownprograms.Open source packages, such as R, that are being developed through a large

knowledgebaseintheonlinecommunity,canhandlelargedatabasesandestimatethemorecomplicatedmodelsneededwhenanalyzingbigdata.

Becauseofthesedevelopmentsweurgefirmsnottolimitthemselvestoonlyoneanalyticalpackage.Thedangerofusingonlyasinglepackageisthatonecannotexploittherichnessofbigdata.Firmsthathavebigdatascientistsintheiranalyticalteamsshouldbeabletofullyexploitthedata—whentheyarenotlimitedbyusingasinglestatisticalpackage.Thedangerhereis,ofcourse,thatanalystsgoforthemostcomplicatedmodelandkeepdiggingintothedataandignorethemanagerialmeaningoftheiranalyses.

Aspecificdisadvantageoftheuseofanalyticalsoftware(eitherstandardoropensource)is that theoutput is frequentlynotuser friendly. Furthermore, themodelsdevelopedarenot always easy for managers to apply. For example, if managers want to make salesforecastsusinganestimatedregressionmodel,theywilltypicallyfail.AsolutionadvocatedbyWhartonprofessorPeteFaderandLondonBusinessSchoolprofessorBruceHardieistousemodelsthatcanbeestimatedandappliedinExcel.Theyspecificallydevelopedforecastmodelswithanunderlyingnegativebinomialdistributionmodel topredictproduct salesandcustomervalue(Fader&Hardie,2001;2009).Theirmodelscanbeextremelyuseful,buttherearenotmanyotherexamplesusingsuchanExcel-basedapproach.

Wehavealreadydiscussedtheroleofmodelsandinsightsinourbigdatavaluecreationmodel.Wewilldiscusstheadditional importantroleofbusinessrulesinthenextsectiononlinkinganalyticstooperationsandspecificallytoactionsandcampaigns.

Analyticalapplications

The fourth layer of big data systems involves analytical applications—that is, reporting,actions and campaigns, decision support, and information-based products and solutions.We have already discussed issues surrounding decision support and information-basedproductsandservices inChapter2.Wewillnowmainly focusonreportingsystemsandactionsandcampaigns.

Reportingsystems

Reportingsystemsaimtomaketheresultsavailabletomanagementthrough,forexample,marketingdashboards.MarketingdashboardshavebecomeratherpopularandaredefinedbyPauwelsetal.(2009:3)as:“Arelativelysmallcollectionofintegratedkeyperformancemetricsandunderlyingperformancedriversthatreflectsbothshortandlong-termintereststobeviewedincommonthroughouttheorganization”.Pauwelsetal.(2009)andReibstein,Norton,JoshiandFarris(2005)proposefivestagesofdashboarddevelopment:

1. Selectingthekeymetrics2. Populatingthedashboardwithdata3. Establishingrelationshipsbetweenthedashboarditems4. Forecastingand“whatif”analysis5. Connectingtofinancialconsequences.

Observingthesesteps,wegenerallyderivetwomainfunctionsofamarketingdashboard.First, it reports theresultsonsomekeymetrics (e.g. retentionratesorNPS)onaregular(e.g.monthly)basis.Second, resultsofmodelsare included insuchawaythatmanagerscan checkwhat occurs if they take specific actions. For example, theymight be keen toknowwhathappenswithcustomerequity if they improve their service (Rust,Lemon,&Zeithaml, 2004).Dashboards are thus filledwithdata and the finalmodel results,whichcanbeusedtoexecute“whatif”analyses.

Thelinkwithoperations:actionsandcampaigns

Figure5.9Linkingdata,analyses,actionsandcampaigns

Inourdiscussiononthedifferentlayersofsystemswemainlytookadataandanalyticalperspective.However,itisimportanttonoticethatoperationalsystemsandprocessesaredirectly linkedwith analytical systems (see Figure 5.9). Operational systems involve, forexample, software that links daily operations (e.g. a call center) to data and specificbusiness rules or selection rules. Business rules and selection rules are the results ofanalytical exercise. These business rules could be that, for example, only customers thathaveahighcustomerlifetimevalue(CLV)—determinedbytheoutcomesofanalytics—getareductiononthepricetheypayforaspecificcontract.Inacallcentertheagentusinganoperationalsystemwillbeabletoobservethecustomerinformationonacomputerscreenand using this system can thenmake offers to customers to renew a contract. Notably,although the interaction occurs in the operational environment, it is sourced back in

databases.Buildingonthecallcenterexample,theinteractionwithacustomerdiscussingtherenewalofacontractcanbeincludedinthedatabase,recordingdataontheoffer,theoutcomeoftheoffer(renewalornot),andsoon.Inabigdataenvironmentthedatacanbeenriched with the unstructured data on the conversation between the agent and thecustomer (e.g. Verhoef,Antonides,&DeHoog, 2004). Selection rules typically involve arule onhow to select customerswhowill receive a specific targeted offer (e.g. an emailcampaignwith a target promotion). These selection rules are very common andused inmanysectors,includingretailingandfinancialservices.Theyaremainlyusedinoutboundcampaigns.

Inadigitalenvironmentoperationalsystemsbecomeveryimportantasinteractionsonwebsitescanbecustomizedbasedonautomatedanalyticalexercises,themselvesbasedondifferentbusinessruleswhichconsiderthebrowsingbehaviorofcustomersatwebsites,theCLVofcustomersetc.Themostfamousexampleofthistechniqueisprobablythatoffirmslike Amazon, which suggest products to customers that have been purchased by othercustomers who have a similar choice behavior (also called “collaborative filtering”).However,incurrentdigitalanalyticalenvironmentsoperationsandanalyticshavebecomeintertwined and the distinction between themhas become blurred.With themillions ofinteractions occurring atwebsites,models can be developed that are constantly updatedwithrelevantnewinformation.Theanalystfirsthastodevelopthe“basic”modelanditsspecific econometric specification.Theestimationparameters can thenbeupdatedbasedontheconstantdigitalinteractionwithcustomersonwebsites,appsetc.,andbasedonthat,constantlyrelevantandtargetedofferscanbeprovidedtocustomersvisitingthesedigitalchannels. In this context Chung, Rust and Wedel, (2009) have developed an “adaptivepersonalization system” and illustrate its implementation for digital audio players. Theproposed system automatically downloads personalized playlists of MP3 songs into aconsumer’s mobile digital audio device and requires little proactive user effort (i.e. noexplicitindicationofpreferencesorratingsforsongs).Theyshowthattheirsystemworksin real time and is scalable to themassive data typically encountered in personalizationapplications.Specifically,theydevelopasystemconsistingof(a)apersonalizedagenda,(b)anadaptivelearningalgorithm,(c)acollaborativecustomizationmodelforallcustomers,and (d) a dynamic customization model for individual customers. For illustration weprovide theirmodel inFigure5.10. Formoredetailswe refer to their paperpublished inMarketingScience (Chung et al., 2009).We expect that due to the development of thesemodelstheanalyticalsystemandtheoperationalsystemswillbecomemoreintegrated.InChapter4.2weprovidedamorein-depthdiscussiononpersonalization.

Figure5.10FlowdiagramoftheadaptivepersonalizationsystemdevelopedbyChung,RustandWedel(2009)

Source:Chungetal.(2009)

Buildingblock4:Organization

Theorganizationalsideofthebigdatadevelopmentgetsrelativelylessattention.Thereisastrongfocusonattractingtalentedbigdataanalystsandtooling.Theorganizationof theanalyticalcompetencewithintheorganizationis,however,importantaswell.Weobservethreespecificchallenges:

1. Centralizationoftheanalyticalfunction2. Cooperationwithotherdepartments/functions3. Presenceofadata-drivenculture.

Centralizationordecentralization

Firmsbuildingupabigdataanalyticalcompetenceface thechallengeofwhere to locatethisfunction.Thisisespeciallyachallengeforfirmsactiveinmultiplebusinessunits.Forexample, firmsoperatingonebusinessunit for theconsumermarketandanother for thecorporate market could have two analytical functions per business unit or one singleanalyticalfunctionservingbothbusinessunits.Firmswithglobaloperationscouldhaveaglobal analytical function serving all separate country organizations or several functionsservinglocalcountryorganizations.Hagenetal.(2013)suggeststhattherearethreemodelsfororganizingananalyticalcompetence(seeFigure5.11).Inthefirstoptionadecentralizedorganizationischosen,whereforeachstrategicbusinessunit(SBU)ananalyticalfunctionissetup.TheadvantagesofthissetuparethattheanalyticalcompetencecanspecificallybedevelopedforeachSBU,servingthespecificneedsofeachSBU.ThefunctionisalsolikelytohaveastrongimpactondecisionmakingineachSBU.Astrongdisadvantage,however,isthatthatfunctionsarerelativelysmallandinefficienciesoccur.SpecificknowledgeandcapabilitiesaredevelopedineachSBUandtherearenoeconomiesofscale.Further,theseseparatefunctionslackanoverallstrategicview.Inthesecondmodel,amoredecentralizedfunctionisdevelopedundertheresponsibilityofoneSBUthatservesthatSBUaswellasother SBUs. This may create more efficiencies and standardize solutions. However, oneproblemisthattherewillbecompetitionamongSBUsfortheanalyticalcapacityanditislikelythattheresponsibleSBUwillbenefitmost. Inafinalmodeltheanalyticalfunctionbecomesanindependentcentralizedstafffunctionservingmultiplebusinessunits.Aswiththe secondmodel, thiswill lead tomore standardizationand less inefficiency.One largepotential challenge with this organization format is that the function can become veryindependent and insufficiently connected with the different marketing departments forimpactful analytical functions. Thiswould indicate a need for analytical functions to bemorecloselylinkedtoaSBU.However,therearesomedisadvantagestothisaswell,asitmay,forexample,leadtoaloweroverallanalyticalskilllevel.

Ingeneralseveralrulescanbeusedtohelpdecidebetweenamoredecentralizedversusamorecentralizedapproach.Decentralizationispreferredovercentralizationwhen:

TherearestronganalyticalskillswithinthefirmThereisalargenumberofanalystsTheanalystteamismatureandindependentIfthereisastrongneedforspecializedknowledgeindifferentSBUsIfteamsdonotdependtoomuchondata-andsoftwaresuppliers.

Cooperationwithotherfunctions

Oneof the problems that analytical functions encounter is the level of cooperation theyhave with other departments and specifically the marketing department. Generally,cooperationbetweendepartmentsisconsideredgood.Thishasbeenshownforcooperationbetweenmarketing & R&D, and betweenmarketing and sales (e.g. Homburg & Jensen,2007;Leenders&Wierenga, 2002). Ithasbeen suggested that somecompetitionbetweendepartments canbegood, as this competitionmay inducea strongerperformanceof theseparatedepartmentsastheyhavetodeliveragoodperformanceinordertogetsufficientsupport fromtheboard.Still, thecompetitionshouldbecombinedwithcooperation.Theterm“coopetition”hasbeencoined for this formof cooperation (Luo,Slotegraaf,&Pan,2006).

Figure5.11Organizationmodelsfortheanalyticalfunction

Source:AdaptedfromHagenetal.(2013)

Another problem is that marketing analysts and, for example, marketing managers,mighthavedifferentthoughtworlds(Verhoef&Pennings,2012),becauseof thedifferenttasks theydoanddifferentpersonalities. Ingeneralanalysts like to focusontheanalysisitselfandarelessinterestedinitsimplications.Moreover,analystswilltypicallyfocusmoreondetailsoftheappliedmethods.Marketingmanagerswouldmovemoreeasilytothenextmarketingaction.

Inspired by thework of the psychologist Carl Jung,we have profiled bothmarketinganalystsandmembersofmarketingdepartmentsonspecificpersonalitytraits (seeFigure5.12). It is interesting to observe that there are indeed strong differences. Analysts areinterestedinin-depthdiscussionsaboutdetailsandstronglyvaluepersonalrelationshipsintheir working environment. In contrast, senior marketing department members andmanagers aremore entrepreneurial and focusmore on outcomes andmaking decisions,takingastrongerleadershiprole.

Figure5.12Differentpersonalityprofilesofanalystsandmarketeers

Thus there are clear differences between analysts andmembers of other departments.Still,formakingsmartmarketingdecisionsstrongcooperationisrequired.Wehavesomesuggestionshowtoachievethis:

Thereshouldbeageneralwillingnessofdepartmentstoworkeffectivelytogether,wherebothcooperationandsomecompetitionoccurs.It is important that both functions understand the different thought worlds andpersonalorientationsoftheother;thismayovercomecommunicationproblemsandleadtomoreeffectivemeetings.Bothdepartmentsshouldhaveasufficientunderstandingofeachother’sexpertiseareas.This implies thatanalystsshouldhavesufficientmarketingknowledge,andmarketingdepartmentmembersshouldunderstandthebasicsofanalytics.Analysts should focusmore onwhat their analytic outcomes of (to be executed)

analysescouldimplyformanagement.

Establishingadata-drivenculture

Amore overarching enabler of a successful implementation of big data analyticswithinfirms is that firms should transform in such away that they relymoreondata in their(marketing-) decision making. This implies a more data-driven culture, which in turnimpliesthatfirmsshouldbasetheirdecisionsmoreonestablisheddata-basedfactsratherthan intuitionorgut feeling.The latter frequentlyoccurswithinmarketing:marketing isfrequentlyblamedforhavingnicecreativeideaswithoutunderstandingtheirimplicationsforbusinessandperformance(Verhoef&Leeflang,2011).AsMcGovern,Court,QuelchandCrawford(2004:74)state:“Themarketingfield ischockablockwithcreative thinkers,yetit’sshortonpeoplewholeantowardananalytic,left-brainapproachtothediscipline.”

Amoredata-drivenculturewithinmarketingisalsofrequentlyreferredtoas“fact-basedmarketing,” clearly suggesting the difference frommore intuition-basedmarketing. Factsaretypicallybasedoninternalanalytics,althoughtheycanalsobebasedonsomegenerallaws in marketing arising from a meta-analysis of executed studies within marketingscience, for example by summarizing a large number of studies on advertising effects.Sethuraman,TellisandBriesch(2011)reportanaverageshort-termadvertisingelasticityof0.12, suggesting that a 10% increase in advertising budgets increase sales by 1.2% in theshort-run.Astrongerfocusonmarketingaccountabilitywithinfirmswilltypicallyinducea stronger data-driven culture. Marketing accountability is generally considered as theextenttowhichmarketingdepartmentsareabletoshowtheeffectsofmarketingactionsonmarketingandbusinessperformancemetricsintheirplans(ex-ante)orevaluations(ex-post)—see for example Verhoef and Leeflang (2009). Firms having a stronger marketingaccountability have more influential marketing departments and also tend to performbetter(Verhoef&Leeflang,2010;O’Sullivan&Abela,2007).

Despite this evidence of the benefits ofmarketing accountability, establishing a data-drivencultureisnoteasy.ThisispartlybecauseofthedifferentthoughtworldsbetweenMIandmarketing.Frequently,marketersdonot sufficientlyembracedata initiativesandsee it as threatening theway they have acted for years.Data and analytics belong to aseparatedepartmentorfunctionnotbelongingtothecoreofmarketing.Oneofthemostfrequently stated comments is that a data-driven culture kills creativity and leads toenduringdiscussionsonthebottom-lineeffects,resultinginmanygoodplansneverbeingimplemented.We believe this comment is based on a wrong view of analytics. In veryspecific tactical tasks (e.g. mail selections, assortment optimization) relying heavily onanalyticswillforsureincreasetheproductivityofmarketing.Thereissufficientevidencetoshowthatusinganalyticsimprovesfirmperformance,forexampleduetothefactthatscarce resourcesarenowallocatedmoreefficientlybetweendifferent customer segments

(e.g. Kumar & Shah, 2009). However, for more strategic long-term decisions, big dataanalytics will play a different role. Here, analytics will provide insights on marketdevelopment,branddevelopments,newunservedsegments,innovationopportunities,etc.,whichshouldbeusedasinputindevelopingmorelong-termorientedstrategicmarketingplans. Model outcomes and insights are then one of the inputs, but also otherconsiderations, includingmanagerial intuition, should be taken into account.As alreadynoted,researchdoessuggestthatacombinationofbothmodelinputsandintuitionleadstothebestdecisions(Blattberg&Hoch,1990).

Onefinalnoteisthatwebelievethatduetotherapiddevelopmentsinbigdata,bigdataareactuallythemselvesadrivingforcebehindmany(service)innovations,whichwerefertoasinformation-basedsolutionsorproducts.Forexample,AmazonfiledforaUSpatentfora“methodandsystemforanticipatorypackageshipping,”whichisbasedonpredictiveanalytics.Thismoveillustratesanincreasingdesireonthepartofmassmarketretailerstodrawonthelatesttechnologiestoanticipateconsumers’needsbeforetheyexpressthemorperhaps are even aware of them.6 Similarly, agricultural Norwegian-based fertilizersupplierYarausesdata-basedsystemstoadjustfertilizingtotheneedsofthecrop.Alsotheweb analytical solutions with adaptive forecasting techniques (Chung et al. 2009)mentioned earlier in the chapter can create innovative service and customer-interfacesolutions.Infactmanycommentatorsintheareaofservicemarketingbelievethatbigdatawillchangehowcustomerswillbeservicedincomingdecades(Rust&Huang,2014).

Topmanagementsupport

Topmanagementsupportisessential,inordertoputinplaceastronganalyticalmarketingfunction that actually has something meaningful to say within the organization; it hasfrequentlybeenshowntobeadriveroftheadoptionanduseofmarketingdecisionsupportsystems and data-based marketing (e.g. Wieringa & Oude Ophuis, 1997; Verhoef &Hoekstra,1999).Actually,whenrequiringchangesinculture,topmanagementsupportisaprerequisite (e.g. Kirca, Jayachandran, & Bearden, 2005). It is required to arrangeinvestments (i.e. acquiring talent, retention of talent, education, and systems) in theanalyticalfunctionsandtoinduceastrongerdata-drivenculture.Althoughthisiswidelyknown,supportforastrongroleofanalyticsdoesnotnaturallyoccur.Thereareacoupleofwaystoachievethis.First,whendiscussingbigdataanalytics,managerscanfocusonthebenefitsofincreasedaccountabilityofmarketing.ThiscanbedonebyarrangingsupportattheCFO level. It is also important to forgea strong connectionbetween the intelligencefunctionand finance.Ourexperience is that this linkcanbepretty successful, especiallywhenthereisagreementonwhichmetricstouse.Second,agreateruseofdatacanfuelagreater focus on the customer. Big data analytics can be used to improve customerexperiences. In today’smarketing environmentweobserve a greater focus on customersandtheirexperienceinmultiplechannels,andanalyticscanbeanimportantingredientfor

successfully delivering this experience (Rust&Huang, 2014). Third, as noted earlier, theincrease in data provides opportunities to create innovations and growth by developinginformation-based solutionsandproducts.Stronganalytical functionscanbepartof thisdevelopmentprocess. So in sum, analytics canhave a strong impact on firms andmanyfunctionsofthefirm(e.g.finance,marketing,R&D).Iftopmanagementisawareofthis,itismorelikelytheywillprovidesupport.

Conclusions

Inthischapterwehavediscussedthebuildingofsuccessfulbigdataanalyticalcapabilitieswithinfirms.Thisisanimportanttopicwhichisfrequentlyneglectedwhendiscussingbigdataopportunities.However,inordertocreatevaluewithbigdata,thesecapabilitiesareofessential importance. First, we considered the internal process for executing impactfulanalyticswithinfirms.Itiscrucialthattheseanalyticsstartwithabusinessquestionorabusinessneed.Analyticsforthesakeofanalyticsdoesnotcreatevalue.Subsequently,wediscussed the important challenges of acquiring and keeping big data analysts, and thedevelopmentofananalyticalfunctionwithinfirms.Weprovidedanin-depthdiscussionofthe different profiles looked for. We considered the systems required for applyingsuccessfulanalytics.Importantly,technologyshouldnotbeleading.Wediscussedthatthebigdataanalyticalsystemwillconsistofdifferentlayers.Finally,weelaboratedonmanyorganizationalissuesconcerningtheorganizationalstructure,cooperationbetweenMIandother functions, the presence of a data-driven culture, and the requirement for topmanagement support. Probablymost important here is that firms require a strong data-drivenculturewhichshouldbesupportedbytopmanagementandshouldbesufficientlyembeddedwithintheorganizationstructuretoensuresuccessfulcooperationbetweentheanalyticalfunctionandother(marketing)functionswithinthefirm.

Notes

1 see www.cisco.com/c/dam/en/us/solutions/enterprise/connected-world-technology-report/Global-Data-CCWTR-

Chapter3-Media-Briefing-Slides.pdf(assessedNovember2,2015).

2ThisexampleisbasedonworkNatashaWalkdidforMIcompany,whereshewasresponsiblefortheMIAcademy.At

thattimePeterVerhoefwasmemberoftheadvisoryboardofMIAcademy.IthasbeenpublishedasacaseinVerhoef

&Lemon(2013).

3Seewww-01.ibm.com/software/data/infosphere/data-scientist/(accessedSeptember27,2015).

4Seeforsomedetailsoneducation:https://nces.ed.gov/pubs2009/2009001.pdf;www.infoplease.com/ipa/A0934035.html

andwww.ara.cat/societat/Linforme-PISA-cinquena-part_ARAFIL20101207_0005.pdf(allaccessedSeptember27,2015)

5 See http://chiefmartec.com/2015/01/marketing-technology-landscape-supergraphic-2015/ (accessed September 27,

2015).

6 See www.atelier.net/en/trends/articles/e-commerce-amazon-hopes-anticipate-consumer-purchases-predictive-

analytics_427319(accessedSeptember28,2015).

References

Blattberg,R.C.,&Hoch,S.J.(1990).Databasemodelsandmanagerialintuition:50%model+50%manager.ManagementScience,36(8),887–899.

Chung,T.S.,Rust,R.T.,&Wedel,M.(2009).Mymobilemusic:Anadaptivepersonalizationsystemfordigitalaudioplayers.MarketingScience,28(1),52–68.

Davenport,T.,&Harris, J. (2007).Competingonanalytics–Thenewscienceofwinning.HarvardBusinessSchoolPress.

Davenport, T., & Patil, D. J. (2012). Data scientist: The sexiest job of the 21st century.HarvardBusinessReview,90(10),70–76.

DeVos,A.,&Meganck,A. (2009).WhatHRmanagersdoversuswhatemployeesvalue:Exploringbothparties’viewsonretentionmanagementfromapsychologicalcontractperspective.PersonnelReview,38(1),45–60.

Fader,P.S.,&Hardie,B.G.S. (2001).ForecastingrepeatsalesatCDNOW:Acasestudy.Interfaces,31(3),94–107.

Fader, P. S., & Hardie, B. G. S. (2009). Probability models for customer-base analysis.JournalofInteractiveMarketing,23(1),61–69.

Germann, F., Lilien, G. L., & Rangaswamy, A. (2012). Performance implications ofdeployingmarketinganalytics. International Journal ofResearch inMarketing,30(2),114–128.

Germann,F.,Lilien,G.,Fiedler,L.,&Kraus,M.(2014).Doretailersbenefitfromdeployingcustomeranalytics?JournalofRetailing,90(4),587–593.

Hagen,C.,Khan,K.,Ciobo,M.,Miller,J.,Wall,D.,Evans,H.,&Yaday,Y.(2013).Bigdataandthecreativedestructionoftoday’sbusinessmodels.HollandManagementReview,148(4),25–37.

Holtom,B.C.,Mitchell,T.R.,Lee,T.W.,&Eberly,M.B. (2008).Turnoverand retentionresearch:Aglanceat thepast, a closer reviewof thepresent, andaventure into thefuture.TheAcademyofManagementAnnals,2(1),231–274.

Homburg, C., & Jensen, O. (2007). The thought worlds of marketing and sales: Whichdifferencesmakeadifference?JournalofMarketing,71(3),124–142.

Humby,C.,Hunt,T.,&Phillips,T.(2008).Scoringpoints:HowTescoiswinningcustomerloyalty.Philadelphia:KoganPagePublishers.

IDG.(2014).BigDataresearch.RetrievedfromIDGEnterprise.RetrievedSeptember11,2015fromwww.idgenterprise.com/report/big-data-2

Kirca, A. H., Jayachandran, S., & Bearden, W. O. (2005). Market orientation: A meta-

analyticreviewandassessmentofItsantecedentsandimpactonperformance.JournalofMarketing,6(2),24–41.

Kumar,V.,&Shah,D. (2009).Expanding theroleofmarketing:Fromcustomerequity tomarketcapitalization.JournalofMarketing,73(6),119–136.

Leeflang, P. S. H., Verhoef, P. C., Dahlström, P., & Freundt, T. (2014). Challenges andsolutionsformarketinginadigitalera.EuropeanManagementJournal,32(1),1–12.

Leenders,M.A.A.M.,&Wierenga,B. (2002).TheeffectivenessofdifferentmechanismsforintegratingmarketingandR&D.JournalofProductInnovationManagement,19(4),305–317.

Luo,X.,Slotegraaf,R.J.,&Pan,X.(2006).Cross-functional“coopetition”:Thesimultaneousroleofcooperationandcompetitionwithinfirms.JournalofMarketing,70(2),67–80.

Manyika,J.,Chui,M.,Brown,B.,BughinJ.,DobbsR.,RoxburghC.,&HungByers,A.(2011).BigDatathenextfrontierforinnovation,competitionandproductivity.RetrievedfromMcKinseyGlobalInstitute.RetrievedOctober,2015fromwww.mckinsey.com/insights/business_technology/big_data_the_next_frontier_for_innovation

McGovern,G.J.,Court,D.,Quelch,J.A.,&Crawford,B.(2004).Bringingcustomersintotheboardroom.HarvardBusinessReview,82(11),70–80.

O’Sullivan,D.,&Abela,A.V.(2007).Marketingperformancemeasurementabilityandfirmperformance.JournalofMarketing,71(2),79–93.

Pauwels,K.H.,Ambler,T.,Clark,B.,LaPointe,P.,Reibstein,D.,Skiera,B.,Wieranga,B.,&Wiesel, T. (2009). Dashboards & marketing: Why, what, how and what research isneeded?JournalofServiceResearch,12(2),175–189.

Reibstein,D.J.,Norton,D.,Joshi,Y.,&Farris,P.(2005).Marketingdashboard:Adecisionsupport system assessing marketing productivity. Marketing Science Conference.Atlanta.

Rigby,D.K.(2014).Digital-physicalmashups.HarvardBusinessReview,92(9),84–92.

Rust, R. T., & Huang, M. H. (2014). The service revolution and the transformation ofmarketingscience.MarketingScience,33(2),206–221.

Rust,R.T.,Lemon,K.N.,&Zeithaml,V.A.(2004).Returnonmarketing:Usingcustomerequitytofocusmarketingstrategy.JournalofMarketing,68(1),109–127.

Sethuraman, R., Tellis, G. J., & Briesch, R. A. (2011). Howwell does advertising work?Generalizations from meta-analysis of brand advertising elasticities. Journal ofMarketingResearch,48(3),457–471.

Tavassoli,N.T.,Sorescu,A.,&Chandy,R.(2014).Employee-basedbrandequity:Whyfirmswithstrongbrandspaytheirexecutivesless.JournalofMarketingResearch,51(6),676–

690.

Verhoef, P. C., &Hoekstra, J. C. (1999). Status of databasemarketing in the Dutch fastmovingconsumergoodsindustry.JournalofMarket-FocusedManagement,3(4),313–331.

Verhoef, P. C., & Leeflang, P. S. H. (2009). Understanding the marketing department’sinfluencewithinthefirm.JournalofMarketing,73(2),14–37.

Verhoef,P.C.,&Leeflang,P.S.H.(2010).Gettingmarketingbackintotheboardroom:Theinfluenceofthemarketingdepartmentincompaniestoday.GfK-MarketingIntelligenceReview,2(1),34–41.

Verhoef,P.C.,&Leeflang,P.S.H. (2011).Accountabilityasamain ingredientofgettingmarketingbackintheboardroom.MarketingReviewSt.Gallen,28(3),26–31.

Verhoef,P.C.,&Lemon,K.N.(2013).Successfulcustomervaluemanagement:Keylessonsandemergingtrends.EuropeanManagementJournal,13(1),1–15.

Verhoef,P.C.,&Pennings,J.M.(2012).ThemarketingfinanceInterface:Anorganizationalperspective.InS.Ganesan,&S.Bharadwaj,(Eds),HandbookofMarketingandFinance(pp.225–243).Cheltenham,UK:EdwardElgarPublishing.

Verhoef,P.C.,Antonides,G.,&DeHoog,A.N.(2004).Serviceencountersasasequenceofeventstheimportanceofpeakexperiences.JournalofServiceResearch,7(1),53–64.

Verhoef,P.C.,Hoekstra,J.C.,&VanderScheer,H.,(2009).Competingonanalytics:StatusquovancustomerintelligenceinNederland.Groningen:CustomerInsightsCenter.

Wierenga, B., & Oude Ophuis, P. A. M. (1997). Marketing decision support systems:Adoption,useandsatisfaction. International JournalofResearch inMarketing, 14(3),275–290.

6Everybusinesshas(big)data;let’susethem

Introduction

Sofar,wehavemainlydiscussedthedifferentbuildingblocksforcreatingvaluewithbigdata analytics. In this chapter we aim to provide real-life examples on how big dataanalyticscanbeimplementedtocreatevalue.Wediscussfivecases,whichvaryinsector,big data approach, and focus. The first case involves mainly traditional analytics, andfocusesonacustomerlifetimevalue(CLV)calculation.Inthenextcases,wemovemoretobig data analytics. InCase 2we present a data integration inwhich data frommultiplesourcesareanalyzedtoimprovemarketingeffectivenessamongmultiplecustomergroups.InCase3wediscussacasethatfocusesonthesystemssideofbigdataanalyticsinordertoachieve more success with personalization efforts. Attribution modeling for an online-retailer isdiscussed inCase4.Our final case focuseson socialnetworkanalytics,whereoneofthekeymessagesisthatsocialhubscanbeidentifiedbylookingatsomeavailablenon-social network variables. When discussing these cases, we apply the storytellingmethodandvisualizationmethodsasdiscussedinChapter4.3,inlinewiththephilosophythatoneshouldpracticewhatonepreaches.

Case1:CLVcalculationforenergycompany1

Situation

The Dutch energymarket has been liberalized and customers can now switch betweenenergy companies. This has resulted in voluntary churn ratesmoving from 0% tomuchhigherlevels.Newproviderswithapricefocushaveenteredthemarket.Asaconsequence,the number of customers of the energy companyhas decreased and revenues are underpressure.Thecompanyrealizesthatcompetitionhasdefinitelychangedandthattheyneedtocompeteforcustomersandcustomervalue.

Complication

Although churn seems to be theproblem, this is not so clear.Customervalue at energycompaniesisnotonlydrivenbychurn,butalsoinvolvesenergyandotherproductusage,servicecosts,payment issuesetc.Thequestion iswhichvaluedrivers they should try toinfluenceandhowtheyshoulddoso.Sofar,thevaluecomponentshavenotbeenstudiedwell. This complication requires a strong conceptualization on the value drivers andgainingtherightdata.

Keymessage

The energy firm should not only look at revenues and retention. When managingcustomersandoptimizingcustomervalue,theyshouldalsoconsiderhowtheycanreduceinboundservicecostsandpaymentenforcementcosts.

Dataandmodelused

The CRM database of the firm was relatively rich and embedded the information onproduct possession, churn, margins, payment method, payment problems, etc.We wereabletoanalyzedatafrom0.9millioncustomers.Beforesettinguptheeconometricmodel,we created a conceptual model on drivers of CLV. Based on this, logit models wereestimatedexplainingeachofthedriversandpredictingtheoccurrenceofanevent.ThesepredictionswereusedtopredicttheCLV.Next,theoutcomesofthemodelswereused,tounderstand how specific drivers can be influenced and which specific actions couldincreaseCLV.

Results

ConceptuallyfourdriversofCLVwereidentified(seeFigure6.1):

Retention(1-churn)Revenues(electricityandgas)Creditlosses(notpaying,paymentenforcementcosts)Servicecosts(includinginboundcalls).

Thepredictionswereusedtounderstandthecontributionofeachofthecomponentsandthisshowedthatrevenuesandservicecostsarethemostimportantvaluedriversintermsof their contribution to CLV. Retention has only a limited contribution (see Figure 6.2),whereas revenues per customer, bad debt (including payment enforcement), and servicecoststronglycontributedtoCLV.

Figure6.1Valuedriversforanenergycompany

Thenextquestion askedwaswhetherone could influence thedrivers.Hence, specificactionswere formulated, togetherwith an idea of their potential success.An immediatereaction is that onewould probably suggest that one should sellmore energy.However,energy usage is largely defined by factors such as household size and the weather.Moreover, from a sustainability perspective energy firms are actually implementingmeasurestolowerenergyusage,forexamplethroughpromotingenergysavinglightbulbs.

Figure6.2ContributionofeachofthevaluedriverstoCLV

Todefinethesuccessprobability,theresearchertalkedwithmanymarketingemployeesto understand which actions would be more likely to be successful. This resulted in aqualitativeassessmentofasuccessprobability.Nextthenumbersofcustomersthatcouldbeaffectedandwhichvaluegaincouldbeachievedwereassessed.Using the simulation

results,itwasfoundthatespeciallyloweringservicecontactsandstimulatingalowerlevelofpaymentenforcementhadthelargestpotentialvalueimpact(seeFigure6.3).

Additionalinsights

The analysis also showed that the value drivers differ significantly per value-segment.Servicecostsandbaddebtbecomeespecially important in the low-valuesegment.Thesecustomerscouldbecomemorevaluableifservicecostswerelower.

Successfactors

The model results helped the company base its marketing actions on customer valuedrivers.Insummary:

Theanalysiswas executed in-housebya seniordataanalyst,whowasverywellawareofscientificstudiesonCLVandwasabletoestimatesophisticatedmodels.The simulation of the model results involves both actual model results andqualitativeassessmentsofsuccess.Therewasa richCRMdatabaseavailable: all the requireddatawereavailable inthatdatabaseanditwasnotnecessarytocollectdatafrommultipledatasilos.Assuchwewereabletoanalyzeaverylargesampleofcustomers.

Figure6.3ImpactofdifferentvaluedriverimprovementsonCLV

Case2:Holisticmarketingapproachbybigdataintegrationataninsurancecompany

Situation

An insurance companywithmultiple brandshas traditionally focusedon a broad targetgroup.Despitethisbroadfocus,thecustomerbaseisnotrepresentativeofthetotalmarketpopulation, being overrepresented with elderly, high-income households. The insurancemarket is in decline and competition is fierce, with a move from traditional high-costchannels(viaanintermediaryorcallcenter)tothelowercostonlinechannels(comparisonsites, direct conversiononwebsite etc.).Competitors are spending substantiallymoreonmediathantheyusedto.Thisputspressureonthemarketshare.Tobetterfacecompetitionand to sustainor (evenbetter) improve themarket share there is aneed to substantiallyincrease the effectiveness of current marketing investments. To realize this, a holisticmarketing approach in combination with a granular perspective of the market andcustomersisdesired.Furthermore,insightsareneededintotheperformancepercustomerprofileonbrand,proposition,distribution,andpricingofthecurrentmarketingmix.

Complication

Inrealizingtheabove,threecomplicationscameupthatmadethisabigchallengefortheorganization. The first complication was organizational: the responsibilities for market,brand,pricing,anddistributionweredealtwithindifferentsiloswithintheorganization,notnecessarilyworkingcloselytogether.Becauseofthis,itwasquiteimpossibletorealizean aligned holistic and consistent marketing approach of the total marketing mix. Thesecond challenge was with the data sources needed to measure and to improveperformance.Thesedatasourceswerecollectedandstoredwiththecustomerintelligencedepartmentandthepricingdepartmentwithintheorganizationandwithseveralexternalmarketresearchagencies.Someofthedatawerenotevenavailableinadatabaseformat,butonlyinhardcopyreports.Becauseofthis,itmaynotcomeasasurprisetolearnthatadataset integrating all these data sourceswasnot available.Consequently, thenecessaryKPIstomeasureandimproveperformancecouldnotbecreatedeasilyduetodifferentdataformats, definitions, aggregation levels, and measure periods/moments. The thirdcomplicationwasthe lackofa frameworkor integralsegmentationthatcouldserveasacommondenominatorforallthedatasourcesandKPIsinscope.Withintheorganizationseveral breakdowns of themarket and the customer basewere available; however, theywerenotaligned.

Keymessage

Asasolutionfor thecomplicationsdescribedabovean interactive,visuallyattractivebigdatadashboardwasdevelopedthatwaseasyforbothmarketersandanalysts touse (seeFigure6.4).Thisenabledashiftfromseparatemeasurements

Figure6.4Thebigdatadashboard

oftheeffectofmarketinginvestmentsandvalueKPIs,todetailedanddynamicmeasuring,interpreting, and forecasting of the marketing performance. In this way, an increase ofperformance and thus return on investment with a granular perspective on customerprofilescouldberealized.

Thesolutionconsistedofseveralelementsthatwerecrucialforsuccess.InthedashboardthecentralelementwasaheatmapthatshowedasegmentationthatwasconsistentforallKPIs in all data sources. The colours in the heatmapwere determined by the over- orunderrepresentationofaspecificKPIforaspecificsegmentintheheatmap.ThedifferentKPIswereclusteredaroundcustomers,brand,andmarket.Furthermore,filterswereaddedtomakeadeepdivepossibleonspecificproductsand/orchannelsand/ortimeframes.

Results

Theresultscreatedbyrealizingthebigdatadashboardapproachwere:

An integrated database, where market data (including competitor performance),brandtrackingdata,customerdata,andcustomersatisfactiondatawereintegratedandpresentedpersegment.Consensus on the set of KPIs to be measured, in order to measure the totalmarketingperformance.A better understanding of totalmarketing performance and its effectiveness, perKPI,persegment,andalsooftherelationshipbetweentheKPIs.Substantial opportunities (multimillion euros) for initiatives to improve themarketingperformanceandmarketingROI.Aroadmapforadditionalrelevantdatasourcestobeadded.

Modelused

Atthestartoftheproject,wedevelopedaconceptualframeworktovisualizetheshifttobemade (seeFigure6.5). In this framework,weshowedthat ideally theeffectivenessof thespendingofthemarketingbudgetwouldbecomevisibleinwhatwecalledtheinputKPIs.However,wealso showed that the inputKPIsare just an intermediary step.Linking theperformance of the input KPIs to the defined output KPIs (measuring V2C and V2Fmetrics), by splitting them out across the segments in the heat map, should make therelationshipsvisible.

Insights

Analyzing the performance using the dashboard in several interactive sessions withmarketers and analysts showed that, for different reasons, different performance levelsaroseinthesegmentsoftheheatmap.Thissuggeststhatdifferentstrategiesarerequiredtoimprove theperformance forspecificcustomerprofiles.Oneof thekey insightswas thattheperformancewith

Figure6.5Theconceptualmodelfortheholisticapproach

youngsterswaslow(lowmarketshare),duetobrandingissues.Atthesametimewesawhigh in- and outflow from this segment, mainly due to aggressive campaigning at thisgroupinacertaintimeframewithonlyonespecificproduct.Theinitiativeaimedatthisgroupwasmore selective targeting (only the potential loyal customers) and at the sametime more focused on offering a broader product range in order to create a moresustainablerelationship.Another insightwasthatduetodelays inupdatesof thepricingmodule(usedformakingacalculationforagoodoffertopotentialnewcustomers)alotofdeals were lost, especially in the family high-income segment, as a result of tooconservativepricingbytheoutdatedpricingmodule.Becausethesegmentatstakeishighvalueandhighvolumeaslightimprovementwouldresultinasignificant(euro)potential.

Successfactors

Thekeytosuccessfortheapproachwasnotjustbyhavingatool—thebigdatadashboard—inplace.Otherfactorsweremuchmoreimportant,suchas:

Using an agile approach, first creating a proof-of-concept within six weeks, andfromtherefurtherimplementingandrefiningthecreatedsolution;Involvingall relevantmarketingdisciplines indefiningtheKPIsand inanalyzingtheresults(i.e.usingmultidisciplinaryteams)Focusingonexecution:howdowetranslatetheinsightsintoaction?Beingpragmatic, startingwithonly thesehigh impactdata sources that couldbeextractedeasilyAggregating thedata sources to thedefineddimensionsof thesegmentation,also

preventingworkingwithtoosensitivedataand/orprivacyissuesMaking the solution scalable and implementable within the organization,preventinghighITimpact.

Case3:Implementationofbigdataanalyticsforrelevantpersonalizationatanonlineretailer

Situation

Theretailerisactiveinagrowingmarketwithmanystrongcompetitors.Toprovidemorevalue to their customers, they aim to inspire customers and provide more relevantrecommendationstotheircustomerswhentheyvisitthecompany’swebsite.Theystrivetogive customers fully automated suggestions of relevant product offers thatmay surprisecustomers. They can already provide customized offers based on some productrecommendationsystems(seeChapter4.2).However,theynowaimtogivemorerelevantpersonalized offers in different settings that reallymake a difference and go beyond the“usualsuspect”offers.

Complication

This retailer hasmillions of customers and offers an assortment ofmore than 8millionstock keeping units (SKUs). Moreover, online customers search, look for, and purchasemultipleproducts,eitheratthesametimeorsequentially.Overallthisleadstoaverylargenumber of customer/product interactions (two billion per year) and even more productrelations, in terms of searches and/or purchases in the same category or multiplecategories,forthesamebrand,forthedifferentthemesoroccasions,etc.Howtoanalyzethese data is not obvious and requires lengthy computation times (i.e. 400 hours).Personalizationbasedonallkindsofproductrelationscantakesolongtocomputethatitmaynot be effective.The challenge is to develop a scalable computationprocess,whereinsteadofmanyweeks,amuchshortertimeperiodisrequired.Thisallowstheretailertocomeupwithmorereal-timeandrelevantoffers,whichshouldleadtohigherconversionrates.

Keymessage

The retailer implementedmap-reducing technology throughwhich,overaperiodof twoyears,thecomputationtimewasreducedtoonedayandtheclickthroughrate(CTR)wasraisedby40%.

Approach

The retailer acquired limited hardware and started to use the open source softwareHadoop.Theytrainedateamofdataexpertstousethesoftware.Theyfirstdidapilotforapre-salesetandusedA/Btesting(seeChapter4.2)totestdifferencesinCTRsbetweentheoldmethodandthenewmethod.Afterthefirstpositiveresultstheyscaleduptoothersetslikesaleandpost-saleandsetsaroundthemesandproductaccessories.

Modelused

To gain product recommendations and other added value propositions from customerinteractionsandbehaviorduringacustomerjourney,analgorithmisneeded.Acustomersearches and/or purchases different products during one or more visits. These productsthereforehavearelationship(seeFigure6.6).Formanycustomersandaproductrangeofover8millionproducts,thiswillresultinbillionsandbillionsofproductrelations.

To make sense of all these data, the retailer developed an algorithm by taking thefollowingsteps(seealsoFigure6.7):

1. Cluster:Thefirststepistorecordallpossibleproductrelationsandtoclusterthesetoequals.

2. Aggregate:Byaggregatingonuniqueproductrelationsthenumberoftimesthateachrelationappearsiscalculated(theproductrelationscore).

Figure6.6Fromsearch/purchasebehaviortoproductcombinations

Figure6.7Algorithmforcalculatingproductrecommendationsbasedontheproductrelationscore

3. Rank:Thenextstepistorankeachproductbyhighesttolowestproductrelationscores.

4. Filter: The last step is to filter out undesirable relations. There are five sorts offilters:

–Noise:relationsthatareveryrare–Policy:unwantedrelationssuchasmedicinesoreroticism–Practical:relationswithproductsthatareoutofstockoroutofrange–Usualsuspects:topfiverecommendations,whicharealreadyrecommended

ingeneral–Derivatives: relationswithproductvariants (for example, silver andgold

iPhones).

Executing thealgorithmoverbillonsofproductrelationsrequiresa lotofcomputer timeandworkingspace.Tomanagethis,theretailerusedMapReduceprogrammingtechnology(seeFigure6.8).TheprinciplesofMapReduceareactuallyquite simple.2MapReducecanprocessalotofdatainashorttimebecauseitsplitsabigtaskintosubtasks.Thesesubtasksare distributed acrossmany computers, which can perform the subtasks simultaneously(distribution). This is done by using the features “map” and “reduce,”which are knownfrom functional programming languages. Results are output files that aremuch smallerthantheinputfiles.Aftersendingthesebacktothecentralserver,thesmalleroutputfilesaremergedintoanaggregatedfinalfile.

Figure6.8MapReduceprogrammingmodel

Results

TheCTRwiththenewmethodwassignificantlyhigherandalmostdoubled.Thisapproachalso led to cost reductions (reducing process time and having an in-house solution).Moreover, ingeneral themethodcreatedmuchmore interestingoffers forcustomers.AnadditionalresultwasgreatercooperationbetweenmarketingandIT,whichmaybehelpfulfordevelopingnewanalyticalprojectsinthefuture(seealsoFigure6.9).

Successfactors

Figure6.9Resultsofnewwayofworking

Thekeystosuccessfortheapproachcanbesummarizedasfollows:

Marketingwasresponsiblefortheproject.Theydid,however,workcloselywithITand also adapted working methods as used by IT (e.g. agile working, scrum

approaches).Thisstimulatedarealcross-functionalapproach.Byusingapilot theanalyst teamwithin the firmcouldshowthat itworkedandtheycouldalsolearnfromtheirmistakes.Fromthepilottheprojectcouldbescaledup.Theorganizationcreatedateamdedicatedtoworkontheprojectthatwasfreetoexperimentwithdifferentsolutions.The software used was free, while the hardware was rather standard and notcomplicated. This substantially reduced the costs, as neither advanced hardwarenorsoftwarehadtobepurchased.Infactthehardwareconsistedoffourstandarddesktops,coupledtocreateamorepowerfulengine.

Insum,thebusinesscasewasrathersimple.Revenuesweresubstantiallyimprovedbythehigher CTRs and conversion rates, and there was a reduction of operational costs, allaccomplishedwithaminimumofinvestment.

Case4:Attributionmodelingatanonlineretailer3

Situation

Anonlineretailersellingmultipleproductswasusingmultipleacquisitiontouchpointsandmedia to attract customers to theirwebsite. Theywanted to improve their allocation ofbudgets over these touchpoints and media. To do so, they needed to know how manycustomers are attracted per touchpoint /mediumand the conversion rate per touchpoint/medium.Theyhaddataonthe lastusedtouchpoint/mediumwhencustomersenterthewebsite, whether these customers made a purchase (conversion), and how large thatpurchasewas(ordersize).

Complication

Themaincomplicationhereismodelingtheeffectivenessofeachtouchpoint/medium.Theretailer traditionallyattributed the sale to the lastused touchpoint /medium.This isalsoknown as the last-click method. That means that if a customer arrives at the websitethroughasearchengineandthensubsequentlybuysaproduct,thevalueofthispurchaseisattributedtothesearchengine.Itisunlikelythatthiseasymethodisgood,ascustomersusemultipletouchpointsandmediaandmightbeinfluencedbymultipletouchpointsandmedia.Inaddition,touchpointsareusedinmultiplephasesofthepurchasefunnel.Hencethecommonbelief is that the last-clickmethod leads to toohighattributionvalues.Thechallengeisthustoachieveabetterattributionmethod.

Keymessage

For this retailerwe developed a new,more accurate attributionmethodwhich could beused to improve the allocation of marketing resources over touchpoints and media. Itturnedoutthatamoresimplemethodthanoriginallyassumedworksaseffectivelyasthemorecomplicatedmodeldevelopedoriginally.

Results

Amoreexactestimateofthetruevalueofatouchpoint/mediumcannowbeassessedthanwaspossibleusingthetraditionallast-clickmethod.Thisespeciallyholdsforsearchengineadvertising, search engine organic search, and direct loads on thewebsite. The value ofemail is underestimated using last-click.An Excel toolwas developed to implement the

newattributionmethodonadailybasis.

Modelused

Wemodeledtheeffectofusedtouchpoint/mediumonconversionandordersize.Onemainissuewithattributionisthatthetouchpoint/mediumusedcanbeendogenous(seeChapter4.2).Tocorrectforthisweusedinstrumentalvariablesandcontrolledforsomebackgroundcustomercharacteristics.Thetouchpointmediumusedinaproductcategorydifferentfromthe one currently visited is used as themain instrumental variable (see Figure6.10). Byincluding customer background characteristics one also can already account for someendogeneity.We thereforealsoestimatedamodelwithout instrumentalvariables,as thismodelislesscomplicatedandeasiertouse.

Insights

The last-click method usually overestimates the value of touchpoints/media. We firstcomparedtheresultsoflast-clickwithamodelaccountingforendogeneity.Ingeneral,thenewmodelresultsinlowerconversionratesandaverageordersizespertouchpoint/media.Onlyforemailareeffectsstrongerinthenewmodel(seeFigure6.11)

We next compared the complicated model with a simpler model that only includedcontrolvariables.Theresultsarerathersimilar,suggestingthatthesimplermodelcouldbeusedaswellandshouldbepreferredbecauseitislesscomplex(seeFigure6.12).

Additionalinsights

Figure6.10Visualizationofmodelbeingused

Thisstudyalsoachievedsomecollateralcatches,forexample:

Thesooneracustomergetsbackonthewebsitethehighertheconversionrate.Thissuggestsstrongopportunitiesforre-targetingofnon-convertedcustomers.

Figure6.11Comparisonofeffectsforattributionmodelandlast-clickmethod

Figure6.12Comparisonofcomplexmodelwithsimplermodel

Priormobilesessionsimprovetheconversionofconsequentwebsessions.(This issomethingwearenowexploringfurther.)ConversionratesarehighestforVIPcustomers.Youngfemalecustomerstendtohaveahigherconversionrate.

Successfactors

Thedevelopmentoftheattributionmodelbenefitedfromthefollowingfactors:

Thefirmwasabletodeliversoliddata.Themodelswere estimated at the individual level instead of the aggregate level,whichwasdoneinapreviousproject.Thisfirmisusedtoworkingattheindividuallevelintheirweb-analyticsandthiswayofmodelingmatchestheirwayofworkingandthinking.There was openness to the use of complicated models, while at the same time

simplermodelswerepreferredifpossible.

Case5:Initialsocialnetworkanalyticsatatelecomprovider4

Situation

Thetelecomprovideroccupiedaleadingpositioninitsmarket.Theyhadstronglyinvestedin a strong analytical function andwere continuously looking forways to extractmorevalue from customers. One of the ways could be to benefit from the social networkscustomersareinandtouse,forexample,viralmarketingcampaignstoincreaseadoptionofnewproductsorreducingchurn.Furthermore,theincreasingrelevanceofsocialmediaatthatpointintimemeantthatmarketingmanagementbecamemoreinterestedinsocialnetworks.Sofar,knowledgeonsocialnetworksisverylimitedbut,giventheavailabledatain which calls between customers are being recorded, a network can be studied andrelevantmetricscanbecalculated.

Complication

Network analysis is really a new field for this company. They have built up extensiveknowledgeoncustomerchurnmodels, lifetimevaluecalculations,responseanalysis,etc.,but the analytics department have no knowledge of network analysis. They thereforedefinedanR&Dprojectwiththeaimofgaininganunderstandingofcustomernetworksandspecificallytounderstandifsocialnetworkmetricscancreateusefulsegments.Theywerealsoworriedabouttheprivacyconsequencesofusingnetworkdata.

Keymessage

Socialnetworkmetricscanbeusedtosegmentcustomersandresultsintoveryinterestingsocialhubsegmentsthatcanbeusedtotargetinviralmarketingcampaignstoincrease,forexample,cross-sellingofnewservicesorserviceimprovementstoreducechurn.

Dataandmodelused

Datawerecollectedoncustomer-to-customerinteractionsusingcalldetailrecords(CDRs).Thesedatashowed thenetworkofcustomersandwith thesedataseveral socialnetworkmetrics can be calculated, such as degree centrality and tie strength (see Chapter 4.2).Thesemetricswerecalculatedpercustomerandsubsequentlya latentclassanalysiswasperformed(seeChapter4.1)tocomeupwithsegments.Thesegmentswereprofiledusinginternal and external data, such as revenue, age (internal CRM database), and

innovativeness(Zipcodelevel).Thisanalysisresultedinfivesegmentsbasedonfitcriteriausedinthesemodels.

Insights

Figure6.13Resultsofclusteranalysisonsocialnetworkvariablesoftelecombrand

Note:aForconfidentialityreasonswedonotprovidetheexactfigures

Five segmentsare found (seeFigure6.13), ofwhich the socialhub seems tobe themostrelevantsegmenttouseforsocialnetworkmarketing.11.6%ofthecustomersbelongtothesocialhubsegment. Interestingly thissegmentalsohasahighAverageRevenue (ARPU),are relatively young and innovative. This implies that by targeting young innovativecustomers with a high ARPU, you can target the social hub customer, that is likely toinfluenceothercustomers.Thisalsoimpliesthatinordertoreachthesehubsnoextensiveclusteranalysishastobedoneusingthenetworkvariables,whichmightcreateproblems,givenprivacyissues.

Successfactors

Afterreflectingonthiscase,weidentifiedtwosuccessfactors:

TheprojectwasreallyanR&Dprojectthatdidnot immediatelyhavetoresult inprofits, and as a consequence an extensive analysis couldbedone inwhichboththeory on social networks and advanced analytics suited for these data could beused.Despite the fact that this projectwas executed in a traineeship, the report had astrong focus onmanagerial outcomes,which increased the impact and perceivedvalueoftheproject.

Conclusions

Wehavediscussedfiveactualcasesonhowdatacanbeanalyzedtocreatevalue.Allthesecases show that firms can actually benefit from analytics. Sometimes these benefits aredirect, but sometimes they are indirect. For example, the online personalization caseimmediatelyresultedinhigherconversionratesonpersonalizedoffers.Inthiscase,therewerealsosomeindirectbenefits,suchastheincreasedcooperationbetweenmarketingandIT.Theenergycompanyexamplecreatednice insights for the firm intowhichactions itcouldpotentiallytaketoincreaseCLV.Intheinsurancecase,therewerealsosomedirectand indirect benefits. Direct benefits included the development of marketing actions toimprovethemarketingperformance.Animportantindirectbenefitisthatthefirmnowhasan overview of all relevant market, brand, and customer metrics and theirinterrelationships.Thekeysuccessfactorsofthesecasesaredifferent,astheprojectsdiffer.However, there seems to be one common success factor, which should explicitly orimplicitlybepresent:Thefirmsshouldhaveadata-drivenculturethatisopentoanalytical(innovative)endeavourstocreatemorevalueforthecustomerandthefirm!

Notes

1ThiscaseisbasedonL’Hoest-Snoeck,VanNieropandVerhoef(2015);formoredetailwerefertothatstudy.Wethank

SietskeL’Hoest-Snoeckforsharinginsightsandsomeinternallyusedpictures.

2ThenameMapReduceoriginallyreferredtoaproprietaryGoogletechnology,butthishassincebeengenericized.A

popularopen-sourceimplementationbasedonthistechnologyisApacheHadoop.

3ThiscasewasjointlyexecutedwithEvertdeHaanandThorstenWiesel.Wekindlythankthemforallowingthiscase

tobeusedinthisbook.

4ThiscaseisbasedonamasterthesisprojectofRicoOoievaar,whichwontheDutchMarketingMasterThesisAward

in2009andwassupervisedbyHansRisseladaandPeterVerhoef.Thisstudywasusedasabasisfor laterworkon

socialinfluenceeffectsaspublishedinRisselada,VerhoefandBijmolt(2014;2015).

References

L’Hoest-Snoeck,S.,VanNierop,E.,&Verhoef,P.C. (2015).Customervaluemodelling inthe energy market and a practical application for marketing decision making.InternationalJournalofElectronicCustomerRelationshipManagement,9(1),1–32.

Risselada,H.,Verhoef,P.C.,&Bijmolt,T.H.A.(2014).Dynamiceffectsofsocialinfluenceand direct marketing on the adoption of high-technology products. Journal ofMarketing,78(2),52–68.

Risselada,H.,Verhoef,P.C.,&Bijmolt,T.H.A.(2015).Indicatorsofopinionleadershipincustomernetworks:Self-reportsanddegreecentrality.MarketingLetters.Forthcoming.

7Concludingthoughtsandkeylearningpoints

Concludingthoughts

Manypeoplehaveclaimed that theuseofbigdata canbeanewgrowthengine foroureconomies.Around theglobe there aremanybelievers in thegreat potential of bigdataand how they can transform companies, marketing strategies, and interactions withcustomers. For us one thing is clear: big datawill changemarketing analytics and howmarketingwillbeexecutedinthecomingdecades.Thiswill,however,notbearevolution,butmoreanevolution.Inrecentdecadeswehavealreadyobservedmanychangesinhowmarketing departments use analytics in their marketing decisions. Marketing decisionshave become more fact-based and market and customer insights have become veryimportant in shaping marketing strategies. Whereas in the 1980s and 1990s marketingintelligence(MI)wasaminorfunction,orabsentaltogetherinmanyfirms,andmarketingscientistswere pushing their developedmodels, we now observe a stronger presence ofanalyticalfunctionsinleadingcompanieslookingforinnovativethougheffectivewaystoanalyzetheirdataandcreateinsights.Theincreasingpresenceoflargechunksofdatawillonlyfuelthisdevelopment.However,itshouldbeclearthatmanagersshouldhavesoundexpectations of these developments. The ultimate goal of marketing analytics andspecificallybigdataanalyticsistocreatevalueforcustomersandthefirm.

In this book, we have aimed to discuss the building blocks of (big) data analytics inmarketing.Followingthetitleofthisbook,webelievethatthisshouldenablemarketingtocreatesmarterdecisions.Thebasisofourdiscussionsisthebigdatavaluecreationmodel,inwhichweshowthatdataasanassetcanbetransformedintopowerfulinsights,smarterdecisionmaking,andinformation-basedproductsorservices,throughbigdatacapabilities.Weexplicitlyhave chosen to take amindful stepback in thebigdatadiscussion,whichsometimesistoomuchahypewiththedangerofbecomingamanagementfad.Buildingonintotalsixdecadesofexperienceinmarketinganalytics,wehavediscussedtheseveralelementsinourbigdatavaluecreationmodel.Importantly,itisourvisionthatasuccessfuluse of (big) data requires (1) good data, (2) strong embedding of the analytical functionwithinafirmhavingtherightsetofcapabilities,(3)strongandimpactfulanalyticalskills,and(4)astrongfocusonvaluecreation.Importantly,thereisnoneedtoexcelatspecificcapabilities.Forexample,itisnotnecessarytorunthemostcomplicatedmodel;infactthatmay be more harmful as it is likely to reduce the impact of the model. Instead usefulinsightscanbegainedwithrelativelysimpleanalyses.

Throughoutthebookwehaveaimedtocombinebothacademicknowledgeandpracticalinsights.Theacademicmarketingliteratureisveryrichintermsofappliedanddeveloped(complicated)modelsformarketinginsightcreationanddecisionmaking.However,thereisfrequentlyagapbetweenacademicsandpractice(e.g.Roberts,Kayandé,&Stremersch,2014).Throughoutourbookithasbeenourobjectivetobridgethisgapbyfocusingfirstonthemanagement issues surrounding the key topics in big data analytics. Second, in the

more in-depth chapters we have aimed to describe more advanced and sometimescomplicatedtopics inamoreaccessibleway.Third, inourapproachweheavilyfocusoncreating impact,byemphasizing the importanceofstorytellingandvisualization.Fourth,wecontinuouslysoughtgoodinsightfulexamplesfromscienceandpracticetovisualizethediscussedmaterial,andweexplicitlydescribedinsightfulcasesinourpenultimatechapter.

Keylearningpoints

After reading all the 13 chapters of this book, we could imagine that a kind of finalsummaryisrequiredonthemainissuescoveredineachone.Weaimtofulfillthisneedinthe final section of this book. It may help the reader to further capitalize on the richknowledgeprovided.ThekeylearningpointsareprovidedinFigure7.1.

Finally, by using text analytics, we have made a word cloud from what we havediscussed inourbook.Thismay furtherhelp ingrasping themost important topics.Weleaveituptothenow“advanced”analysttointerpretthiscloudprovidedinFigure7.2.

Figure7.1Keylearningpointsbychapter

Figure7.2Wordcloudofourbook

Reference

Roberts,J.H.,Kayandé,U.,&Stremersch,S.(2014).Fromacademicresearchtomarketingpractice: Exploring the marketing science value chain. International Journal ofResearchinMarketing,3(2),127–140.

Index

3“V”sofbigdata15,75,132

5“W”smodel85–87

Aaker,David33

A/Btesting198–199,294

Abela,Andrew247

Achrol,R.S.221

acquisitionofanalysts263,266–267

adaptiveforecastingtechniques215–216

adaptivepersonalizationsystem275–276

adoptionmodel26

“adversebehavior”metrics58

advertisingawarenessmetrics28

affectivecommitment39

affectivemeasures28

aggregationofdata112,114,185,206

Ailawadi,K.L.44,54–55

Allenby,Greg211

analysisstrategies(bigdata)122–127

analysts,profileof263–264,279

analyticalapplications(analyticalcompetencesystems)269,270,273–276

analyticalchallenges(ofdataintegration)100,101,103–104

analyticaldataplatform(analyticalcompetencesystems)269,270,271–273

ANCOVA(analysisofcovariance)199

anonymizingdata113,114

ANOVA(analysisofvariance)199

Ansari,A.208

Anscombe,Francis238–239

APE(averagepredictionerror)166

appendeddata96–97

APS(adaptivepersonalizationsystems)210

ARPU(Averagerevenueperuser)303

Ataman,M.B.216

ATL(abovetheline)campaigns64,71

attributeanalysis(classicdataanalyticstechnique)143,172–180

attributionmodeling195,203–206,298–301

B2B(business-to-business)markets2,77,88

B2C(business-to-customer)markets2

“bagging”135,185

“bagging-boosting”technique135

barcharts243,246

Bass,F.M.222

BAV(Brand-AssetValuator®)29–30,31,32

Bayesianmodels210–211,216

BE(brandequity)20–21,32–33,40,53–55,267

behavioraltargeting14,17,125

Berger,P.D.65

betweennesscentrality223–225

bigdataanalytics(bigdatavaluecreationmodel)9,13–16

bigdataassets(bigdatavaluecreationmodel)9–10

bigdatacapabilities(bigdatavaluecreationmodel)9,10–13

bigdatadashboard289–291

“bigdatahubris”212

bigdatavalue(bigdatavaluecreationmodel)9,16–21

bigdatavaluecreationmodel9–22,305

big-data-basedsolutions17

Blattberg,R.C.184

Bolton,R.N.36,41

BostonConsultingGroup107

branddata(datasource)82–84

brandequitypreference-basedapproach54

brandequitypricepremiumapproach54

brandequityshareholdervalueapproach53

brandfunnels82

brandlevelchanges(bigdataanalytics)127,128,130–131

brandmetrics:brandawarenessmetrics28,29;brandconsiderationmetrics28–29;brandevaluationmetrics53–55;brand

likingmetrics28,29;brandloyaltymetrics52–53;brandmarketperformancemetrics51–53;brandpenetrationmetrics

51–52;brandpreferencemetrics28–29,30;brandsalesmetrics51–52;V2Cmetrics27–35,36;V2Fmetrics29,32, 33,

51–55

brandrepurchaserates52–53

brand/productlevelanalysis4–5

Briesch,R.A.280

BTL(belowtheline)campaigns64

bubblecharts241–242,246

Bügel,M.S.39

buildinganalyticalcompetences:challengesto253–254;organization254,276–282;organizationaltransformation255–259;

people254,263–268;process254,259–263;systems254,268–276

bulletcharts243,244

businesschallenges(ofdataintegration)100,101,104

businessquestions259–262

Buunk,A.P.39

C2C(customer-to-customer)markets2,222

calculativecommitment39

case studies 285; attribution modeling 298–301; CLV calculation 286–288; holistic marketing approach 289–293;

personalizationthroughbigdataanalytics293–298;socialnetworkanalytics301–303

CDRs(calldetailrecords)223,302

centralization/decentralizationoforganizations276–277,278

CES(customereffortscore)36,37–38

CFMs(customerfeedbackmetrics)35–38,134–135,214–215

CHAID(Chi-squareautomaticinteractiondetection)

analysis181–182,183

channelswitching70

channelusagedata199–203,205,207

Chen,H.217

Chi-squaretests145–146,184

choice-basedconjointanalysis176–178

Chung,T.S.210,275–276

circularnetworkcharts241–242

CIV(customerinfluencevalue)68–69,225

CKV(customerknowledgevalue)68–69

classicdataanalytics:attributeanalysis143,172–180;customersegmentation142,155–162;historyofmarketinganalytics

140–141;migrationanalysis142,150–155;predictivemodeling143,180–189;profiling142,145–150;reporting141–142,

144–145;trendanalysismarketandsalesforecasting142,163–172

clickstreamdata194

CLM(closedloopmarketing)206,207

closenesscentrality223–225

clusteranalysis155–157,159–162,178

CLV(customerlifetimevalue):approach/objectivesofstudy5;andbigdataassets10;calculating64–66;componentsof

59–64;andcustomerequity67,68;anddataintegration99;defining58–59;driversof286–288;energycompanycase

study 286–288; andmarketing ROI 70–71;model building 66–67; and operational-analytical linkages 274–275; and

predictiveanalyses122;andV2Fmetrics49,58–69,70–71

CMOs(chiefmarketingofficers)1

cognitivebrandmetrics28

COGS(costsofgoodssold)60–61

cohortanalysis154,155

collaborativefiltering207–208,275

collateralcatch123,126–127

columncharts243,246

columnhistograms245,246

combiningaggregationlevels213–214

commercialdataenvironment76–79,87–88,93,95,99–100,101–102

commitment39–40

comparisoncharts242–243,246

competitiveintelligencedata77–78

completenessofdata(dataqualitydimension)89–90

complication(storytellingcomponent)235,237,238

compositioncharts243–245,246

computersciencemodels135–136,185

conjointanalysis54,140,174–180

contentfilteringsystems207–208

continuousdatacollection134–135

contractuallifetimesetting63

conversionratemetrics56

“coopetition”278

Coopmans,Baptiest257–258

corporatereputation43

“cosmopolitans”(marketdatacategory)80

costallocationoptions61–62

Court,D.280

CPC(costperclick)70

CPM(costpermille)70

CPO(costperorder/transaction)70

Crawford,B.280

creditlosses(CLVdriver)286–288

creditworthiness77

CRM(customerrelationshipmanagement):andanalyticalcompetencesystems270–271;andapproach/objectivesofstudy

4–5;andclassicdataanalytics141;andcustomerlifetimevalue58;developmentof1,8;dominantroleof IT3; and

internaldatasourcesofmetrics41–42;andimportanceofanalytics118,131

cross-buyingrates57

Croux,C.185

CRV(customerreferralvalue)68–69,225

CSR(corporatesocialresponsibility)42,43

CTR(clickthroughrate)294,297,298

culture,roleof12–13

customercrossings145–146,150

customerdata(datasource)84–85

customerdescriptors96

customerequity40–41,67,68

customerfeedbackloop35–36

customerheterogeneitymodeling210–211

customeriddata84,85,88,89,101,102,271

“customerintimacy”39,40

customerjourneyanalysis195,199–203

customerlevelanalysis4–5

customerlevelchanges(bigdataanalytics)127,128,131–132

customermetrics: customer acquisitionmetrics 55,56; customerdevelopmentmetrics 55, 56–58; customer engagement

metrics67–69;customerjourneymetrics69–70;customersatisfactionmetrics36,37–39;customervaluemetrics55,58;

V2Cmetrics35–42;V2Fmetrics55–70

customerprivacy:bigdataas threat to3,11,106–107;anddatasecurity114–116;data storageandusage105–106, 108,

111–112; defining privacy 107; and ethics 110–111; government legislation 108–110; and internal data analytics

112–113;privacyconcerns106,107–108,112;privacypolicies111–112,113

customersegmentation(classicdataanalytics

technique)142,155–162

customertrust39–40

customer-specificdata41

CVM(customervaluemanagement)78,88,90,181,184

databeinguptodate(dataqualitydimension)89–90

datadisappointment8

dataenthusiasm8

datafusion87,91,135,213,271

dataintegration:analyticalchallenges100,101,103–104;businesschallenges100,101,104; inbigdataera100–104; and

data types 95–99; and data warehouses 100, 101; individual level integration 102; integrating data sources 93–95;

intermediatelevelintegration102–103;technicalchallenges100,101–103;timelevelintegration103

datamarts271–272

datamining123,125–126

datamodeling123,124–125

dataprotectionregulations109–110

dataquality89–90

datarealism8

datascientists263,268,272–273

datasecurity114–116

data sources: 5 “W”smodel 85–87; and analytical competence systems 269, 270–271; brand data 82–84; customer data

84–85;anddataintegration93–104;anddatawarehouses87–88;external76–78;andholisticmarketingapproach289,

291,293;and integratedbigdatamodels212–213; internal76,78–79;marketdata80–82;multi-sourcedataanalysis

127,130,131,196;structured76,79–80;unstructured76,79–80;useof85–87

datastoring(analyticalcompetencesystems)269,270,271

datawarehouses87–88,100,101,271

databasestructures88–89

data-drivenculture280–281

Davenport,T.119–120

DeBruin,B.110–111

DeHaan,E.206

DeVries,L.13,16

decileanalysis146–147,148

decisiontrees181–182,183

declareddata96

degreecentrality223–234

Dekimpe,M.G.170

demand-sidebranddata82–83

demand-sidecustomerdata84

demand-sidemarketdata80–81

dendrograms160

departmentalcooperation277–280

descriptiveanalyses144–145,221,261,272

digitalbrandassociationnetworks33,34

“digitalidentity”107

digitalinformationoverload231–233

digitalsummaryindices33–35

digitalizationofsociety1,2–3

directmarketing141,180,206

discretechoicemodels63

distributioncharts245–247,246

Dixon,M.36,38

Donkers,A.C.68

Donkers,B.65,91,122,136

doublebarcharts245

DSI(digitalsentimentindex)35,271

durationmodels63

DutchNationalBank106

DutchTelco(telecommunicationscompany)18

dynamicanalyses121–122

dynamictargeting196,206–211

Eberly,M.B.267

EBITDA(earningsbeforeinterest,taxes,depreciationandamortization)59–62

Eggers,F.176,177

Ehrenberg,Andrew53

Eliashberg,J.225

Essegaier,S.208

ethicaldecisionmaking110–111

ETL(extraction,transformationandloading)process93–95,271

eWOM(electronicword-of-mouth)34,42

expectedlifetime(CLVcomponent)63

ExperianUK(externaldataprovider)97–98

explanatoryanalyses121–122

explosionofdata1–3

externaldatasources76–78

externalprofilinganalyses147

extractionstage(ETLprocess)94

Facebook35,36

Fader,Pete273

failurefactorsforcreatingimpact233–234

Farris,P.20,72,273

featuregeneration219–220

Feld,S.56

Few,Stephen243

flupredictiondata212

FMCGs(fast-movingconsumergoods)127,130

Fombrun,C.J.43

Franses,P.H.184

Frenzen,H.56

FTC(FederalTradeCommittee)110

F-tests145–146

fullpopulationanalysis132–133

“fundamentalist”(privacyconcernsegmentation)108

fuzzymatching219

geographicmaps241–242

Gijsenberg,M.J.114

Ginicoefficient186–188

GMOK(generalizedmixtureofKalmanfilters)model114

Goodwin,C.107

Google(internetmultinational)27,51

GRPs(grossratingpoints)53

Hadoop(software)11,76,270,271,294

Hagen,C.277

Hanssens,D.M.32,170

Hardie,Bruce273

Harris,J.119–120

hierarchicalclusteranalysis155,159,162

Hinz,O.225

hitrate184,186

Hoekstra,J.C.99

holisticmarketingapproach289–293

Holtom,B.C.267

Holtrop,N.114

homophily225

Hu,M.217–218,219

Hunneman,A.172

hype8

idiosyncraticmodels13

implieddata96,98–99

individualleveldataintegration102

infographics232,249

InformationisBeautiful(book)249

infrequentobservations188

ING(bank)3,106

insights(bigdataanalytics)13,16

instrumentalvariables168,299

integratedbigdatamodels196,212–216

intermediateleveldataintegration102–103

internaldatasources41–42,76,78–79

invoicedata78,79

Jacobson,R.32

Jones,K.106

Joshi,Y.273

Jung,Carl279

Kamakura,W.A.91,161

Keller,KevinLane32

Kennedy,R.212

keylearningpoints306–307

King,G.212

K-meansclusteranalysis155,156,159–160,162

KNABBank14

Kohli,R.208

KonuşU.161–162,200,202

Kotler,P.140,221

Koyckmodel168

KPIs (key performance indicators): and building analytical competences 261–262, 263; and data integration 104; and

holisticmarketingapproach289–293;andlike-4-likeanalysis153;andmigrationmatrices152;reporting141,144

KPN(telecommunicationscompany)18,256,257–259

Krafft,M.56

Kumar,V.68

L4L(like-4-like)analysis153–154

laggedvariables128,169

“last-click”method203,205,298–300

latentclassanalysis161,178,200,202,208,210,302

Lattin,J.M.155

Lazer,D.212

Lee,T.W.267

Leeflang,P.S.H.1,13,99,140,165,168,280

Lehmann,D.R.54–55

Lemmens,A.136,185

Lemon,K.N.19,36,40–41

liftcharts245

Lilien,Gary140

linecharts243,245,246

linearregressionmodels163–166

Liu,B.217–218,219

loadingstage(ETLprocess)95

logitregressionmodels182–185

logit-modelanalysis63

log-logtransformation164

Malhotra,N.K.146

managementsupport262–263,281–282

MAPE(meanabsolutepercentageerror)166

MapReduceprogrammingmodel76,296

margin(CLVcomponent)59–62

marketdata(datasource)80–82

marketlevelanalysis4–5

marketlevelchanges(bigdataanalytics)127,128–130

marketmetrics:market attractivenessmetrics 50–51;market sharemetrics 51–52, 53; V2Cmetrics 26–27; V2Fmetrics

50–51

marketingaccountability280

marketingcampaigns16–17

marketingchallenges123

marketingdashboards273–274

marketinggrowthobjectives123

marketingresearchdata77

marketingtechnologyvendors268–269

Markovmodels154–155

McCandless,David249

McDonalds(restaurantchain)33,34

McGovern,G.J.280

McKinseyGlobalInstitute1

MCMC(MarkovChainMonteCarlo)method210,211

Meer,David8

Mela,C.F.216

Merkel,Angela106

message(storytellingcomponent)235,237–238

MI(marketingintelligence)255–259,260,262–265,268,271,280,305

migrationanalysis(classicdataanalyticstechnique)142,150–155

migrationmatrices152

Miller,GeorgeA.235–236

Miller’slaw235–236

“mind-setmetrics”32

Minto,Barbara234

missingdatavalues91

Mitchell,T.R.267

Mizik,N.32

mobiledata79–80,87,226

modeldevelopment(bigdataanalytics)13–14

moralintensity110–111

“moreadvanced”analysismodels154–155

“morein-depthanalysis”154

Motivaction(researchagency)80–81

MSE(meansquarederror)166

multi-channelusage161–162

multi-facetedcustomerbehavior188–189

multi-levelmodel214–215

multi-sourcedataanalysis127,130,131,196

Naert,P.A.140

Nasr,N.I.65

NBD(negativebinomialdistribution)models63

Neslin,S.A.54–55,90,183,200,202

networkcharts241–242

networkvaluation225

Netzer,O.155

“neuralnetworks”135

newproductsalesmetrics51

NLP(naturallanguageprocessing)217

nomistakesindata(dataqualitydimension)89–90

Nokia(telecommunicationscompany)19

non-contractuallifetimesetting63

non-linearadditivemodels164

non-stationaryvariables169–170

normativecommitment39

Norton,D.273

NPS(netpromoterscore)36,37–38,44,98,124,134,144

NSA(NationalSecurityAgency)3,105

“numberoflevelseffect”(conjointanalysis)176

OLAP(onlineanalyticalprocessing)120

OLS(ordinaryleastsquares)regressionmodels164

onetimeinvestments(CLVcomponent)63–64

opensourcesoftware11,272–273,294

operational-analyticallinkages274–275

opinionanalysis217–221

opinionleadership223

“opinionmining”217

opinionwordextraction/assessment220

organization(analyticalcompetencebuildingblock)254,276–282

organization(bigdatacapability)12

overheadcosts60

overlaiddata96,97–98

oversampling188

Paap,R.184

“pathtopurchase”models69–70,203–204

Pauwels,K.32,273

“paymentequity”metric40

PCA(principalcomponentsanalysis)149,157–159

people(analyticalcompetencebuildingblock)254,263–268

people(bigdatacapability)10

people(datasecurityelement)115

permission-basedmarketingapproach112,113,114

personalizationsystems206,208,210–211

Peters,K.56

piecharts244,246

Podesta,John106

Polman,Paul42

“polygamousloyalty”52

POST(part-of-speech-tagging)217,218

“post-materialists”(marketdatacategory)80

powerofanalytics119

“pragmatist”(privacyconcernsegmentation)108

pre-attentiveattributes247–248

pre-defineddata38,123,126

predictiveanalyses121–122

predictivemodeling(classicdataanalyticstechnique)143,180–189

predictiveperformancemeasures185–188

pricefairnessmetric40

pricepremiummetrics54

problemsolving123,124

process(analyticalcompetencebuildingblock)254,259–263

processes(bigdatacapability)11–12

processes(datasecurityelement)115,116

productrelationscore294,295

profiling(classicdataanalyticstechnique)142,145–150

pruning219–220

pseudomyzingdata113,114

PSQ(perceivedservicequality)171

purchasefunnels126,194,198,203–204,298

pyramidprinciple234–235

Quelch,J.A.280

R(opensourcesoftware)11

R&D(researchanddevelopment)13,53,278,301,303

radarcharts244

“rangeeffect”(conjointanalysis)176

ranking-basedconjointanalysis176

rating-basedconjointanalysis176,178

RE(relationshipequity)41

real-timemodels136–137

recommendationsystems206,207–208,209

“regimeswitching”models216

regression-basedapproachattributeanalysis173–174

regressionmodels163–166,167–168,169–172,182–185

Reibstein,D.J.273

Reichheld,F.F.38

relationshipcharts241–242,246

relationshipcostsandriskmetrics57–58

relationshipexpansionmetrics57

relationshiplengthmetrics56–57

“relationshiplifecycle”55

repetitionofmeasures/respondents134–135

reporting(classicdataanalyticstechnique)141–142,144–145

reportingsystems(analyticalcompetencesystem)273–274

RepTrack®measurementsystem43

ReputationInstitute43

researchshoppingpercentage70

responseratemetrics56

retention(CLVdriver)286–288

retentionofanalysts263,267–268

revenuepremiummetrics54–55

revenues(CLVdriver)286–288

“reviewvolume”42

“reviewsvalence”42

RFM(recency,frequencyandmonetaryvalue)model180–181,182

Risselada,H.136

Rogers,E.M.26

ROI(returnoninvestment)14,70–71

Roose,Marteyn258

Rosling,Hans251

Rossi,Peter211

Rust,R.T.19,40–41,67,210,275–276

Sattler,H.176,177

SBUs(strategicbusinessunits)127,277,278

scannerdata118

scanningrevolution8

SCANPROmodel140

scattercharts241–242,246,247

search/purchasebehavior294,295

seasonaleffects167,168

security11

selectionofanalysts263–266

“self-hiddenEastereggs”189

sentimentanalysis217–218,220,221

SEO(searchengineoptimization)198

servicecosts(CLVdriver)286–288

Sethuraman,R.280

Shah,D.58

“shareofheart”metrics25

“shareofmind”metrics25

Sharp,Byron53

Shepard,David140–141

showroomingbehavior200,201

site-centricclickstreamdata194

situation(storytellingcomponent)235,236–237,238

sizeofeffects133–134

SKUs(stockkeepingunits)294

Slice(app)82

Sloot,L.M.172

Snowden,Edward3,105

socialdatacollection218–219

sociallistening197,216–221

socialmedia:andbrandmetrics33,35,36;andcustomerengagement67–69;andexplosionofdata2–3;andexternaldata

sources78,79;andsocial listening197,216–221;andsocialnetworkanalysis197,221–225; social network analytics

casestudy301–303;andsocialtargeting225;socialnetworkdata222–223;socialnetworkmetrics223–225

socialnetworkanalysis197,221–225

socialnetworkanalyticscasestudy301–303

socialnetworkdata222–223

socialnetworkmetrics223–225

socialtargeting225

“softdata”80

sophisticationlevelsofanalytics120–121

Spring,P.N.99

SQL(structuredquerylanguage)queries271

Srinivasan,S.32

Srinivasan,V.155

stackedcharts244–245,246

standardizedmodels13–14

starcharts243,244

statedimportancemeasurement172–173

staticanalyses121–122

Sternberg,R.J.39

storytelling:anddigitalinformationoverload231–233;clearstorylinechecklist236–238;coremessage234,235,236,237;

importanceof231–232;integrationwithvisualization250,251;pyramidprinciple234–235

structureddatasources76,79–80

substantiveeffects133–134

summarygeneration221

supply-sidebranddata82–83

supply-sidecustomerdata84–85

supply-sidemarketdata80–81

survivalanalysis154,156

“switchingmatrix”52,67

system(datasecurityelement)115

systems(analyticalcompetencebuildingblock)254,268–276

systems(bigdatacapability)11

tagclouds244

TAM(technologyacceptancemodel)26

teamapproach265–266

technicalchallenges(ofdataintegration)100,101–103

Tellis,G.J.280

Tesco(retailer)14,123,126

textanalytics217–221,226,306

Thaler,Rich14

tiestrength225

Tillmanns,S.68

timeleveldataintegration103

timeseriesdata/models42,122,127–129,142–145,163,166,168–172,215

“top-2-box”customersatisfaction37–38

top-decilelift186,187

touchpointdata199–201,203,205–206

transactiondata96–97

transformationstage(ETLprocess)95

trendanalysismarket/salesforecasting(classicdataanalyticstechnique)142,163–172

trust39–40

Tucker,Catherine112,113

UGC(usergeneratedcontent)33

“unconcerned”(privacyconcernsegmentation)108

“unpooledanalysis”172

unstructureddatasources76,79–80

user-centricclickstreamdata194

V2C (value to the customer) metrics 20–21, 25; balance between V2C and V2F 17–19; brand metrics 27–35, 36; and

continuousdatacollection134–135; customermetrics35–42; defining 17, 22; limitingmetrics collection 44;market

metrics26–27;andV2Smetrics25,42–43,45

V2F (value to the firm)metrics 20–21, 49; balance betweenV2C andV2F 17–19; brandmetrics 29, 32, 33, 51–55; and

customerlifetimevalue49,58–69,70–71;customermetrics55–70;defining17,22;marketmetrics50–51;marketing

ROI70–71

V2S(valuetosociety)metrics19,25,42–43,45

valencescores34

validationprocess94

valuecreationconcepts17

“valuedelivery”17

“valueextraction”17

valueofdata15,132

VanBruggen,G.225

VanderLans,R.225

VanHeerde,H.J.216

VanRiel,C.B.M.43

Vanhuele,M.32

VAR(vector-autoregressive)models170–171,206

varietyofdata15,75,132

VE(valueequity)40

velocityofdata15,132

Venkatesan,R.68

veracityofdata15,132

Verhoef,P.C.1,36,39,56,91,99,114,136,172,200,202,280

Vespignani,A.212

VirginMobile(telecommunicationscompany)18

visualization:Anscombe’sQuartet238–239;choosingcharttype241–247;anddigitalinformationoverload231–233;graph

design247–249;importanceof231–232;integrationwithstoryline250,251;objectivesofvisualizingdata240;powerof

16;trendsin249,251

volumeofdata15,132

Waldstatistic184

Ware,Colin247–248

waterfallcharts244,246

webanalytics194,195,198–199

webroomingbehavior200,201

Wedel,M.91,161,210,275–276

WhatsApp(app)129–130

Wierenga,B.225

Wieringa,J.E.114

Wiesel,T.68

“willingnesstopay”54

win-backmetrics57

wordclouds219,244,249,306,308

Zeithaml,V.A.19,36,40–41

zipcodedata/analysis77,78,79,97–98,148,149,150,222