luo’s qes miguel alvarez javed jussa quan?ta?ve …com.estimize.public.s3.amazonaws.com/l2q...

29
Yin Luo, CFA Vice Chairman Quan.ta.ve Research, Economics, and Por8olio Strategy [email protected] Miguel Alvarez [email protected] Javed Jussa [email protected] Sheng Wang [email protected] Gaurav Rohal, CFA [email protected] QES Desk Phone: 1.646.582.9230 [email protected] Luo’s QES Quan?ta?ve Research, Economics, and PorDolio Strategy October 2017 DO NOT FORWARD – DO NOT DISTRIBUTE – DOCUMENT CAN ONLY BE PRINTED TWICE This report is limited solely for the use of clients of Wolfe Research. Please refer to the DISCLOSURE SECTION located at the end of this report for Analyst Cer.fica.ons and Other Disclosures. For Important Disclosures, please go to www.WolfeResearch.com/Disclosures or write to us at Wolfe Research, LLC, 420 Lexington Avenue, Suite 648, New York, NY 10170

Upload: phungnhan

Post on 17-May-2018

216 views

Category:

Documents


2 download

TRANSCRIPT

Page 1: Luo’s QES Miguel Alvarez Javed Jussa Quan?ta?ve …com.estimize.public.s3.amazonaws.com/l2q presentations/3_l2q_ldn... · We inves.gate a wide range of machine learning algorithms

YinLuo,CFAViceChairmanQuan.ta.veResearch,Economics,[email protected]@wolferesearch.comJavedJussaJJussa@[email protected],[email protected]:[email protected]

Luo’sQES

Quan?ta?veResearch,Economics,andPorDolioStrategy

October2017

DONOTFORWARD–DONOTDISTRIBUTE–DOCUMENTCANONLYBEPRINTEDTWICEThisreportislimitedsolelyfortheuseofclientsofWolfeResearch.PleaserefertotheDISCLOSURESECTIONlocatedattheendofthisreportforAnalystCer.fica.onsandOtherDisclosures.ForImportantDisclosures,pleasegotowww.WolfeResearch.com/DisclosuresorwritetousatWolfeResearch,LLC,420LexingtonAvenue,Suite648,NewYork,NY10170

Page 2: Luo’s QES Miguel Alvarez Javed Jussa Quan?ta?ve …com.estimize.public.s3.amazonaws.com/l2q presentations/3_l2q_ldn... · We inves.gate a wide range of machine learning algorithms

•  Targe?ngMergersandAcquisi?ons.Inatypicaltakeoverdeal,theacquireroffersasignificantpremiumoverthetargetfirm’sprevailingshareprice.Theupsidefromaccuratelyiden.fyingtakeovertargetsissignificant.

•  Alterna?veBigDataFactors.Inaddi.ontotradi.onalstock-selec.onfactors,weconstructasetofunconven.onalfactors.Inpar.cular,wefindcorporatefilings(e.g.,Forms8-K,4,13D/13G,DFAN14A,etc.)fromtheEDGARdatabasecontainsusefulinforma.onaboutfutureM&Atransac.ons.UsingtheTRSDCdatabase,wealsoconstructpastM&AeventcountsignalsandourproprietarySectorM&AMomentumfactor.

•  MachineLearningTakeovers.Todealwiththerareunbalancedsamplenatureoftakeoverpredic.on,weintroducematchedsampleandbaggingtechniques.Weinves.gateawiderangeofmachinelearningalgorithmsintakeoverpredic.on:CART,randomforest,SVM,ANN,andAdaBoost.OurfinalSMAP(Systema.cMergerandAcquisi.onPredic.on)modeltakesadvantageofforecastcombina.on,byblendingsevenmachinelearningmodels.TheSMAPmodeldemonstratesexcellentpredic.vepowerintakeoverpredic.on.

•  Applica?onoftheSMAPModel.Discre.onarymanagerscanusetheSAMPmodelasapre-screeningtoolbeforeconduc.ngmorein-depthfundamentalresearch.Buyingthosestockswiththehighesttakeoverprobabili.esachievessuperiorreturn.Long/shortmanagerscanusetheSMAPtoavoidshor.ngpoten.altakeovertargets.TheSMAPoverlayontheshortsidenotonlyboostsreturn,butalsoreducepor8oliovola.lity.

2

10.MachineLearningTakeovers

Page 3: Luo’s QES Miguel Alvarez Javed Jussa Quan?ta?ve …com.estimize.public.s3.amazonaws.com/l2q presentations/3_l2q_ldn... · We inves.gate a wide range of machine learning algorithms

3Sources:FTSERussell,S&PCapitalIQ,ThomsonReuters,WolfeResearchLuo’sQES

a)OverviewofM&ATarget

EventTypeFrequency ExcessReturnAroundtheEventDate

EventDefini?onNumberofM&AAnnouncements

Page 4: Luo’s QES Miguel Alvarez Javed Jussa Quan?ta?ve …com.estimize.public.s3.amazonaws.com/l2q presentations/3_l2q_ldn... · We inves.gate a wide range of machine learning algorithms

4

Characteris?csofM&ATargets

AverageCumula?veExcessReturns MedianCumula?veExcessReturns

FactorExposureBeforeandAXertheM&AAnnouncement AdjustedM&AFrequencybySector

Sources:FTSERussell,S&PCapitalIQ,ThomsonReuters,WolfeResearchLuo’sQES

Page 5: Luo’s QES Miguel Alvarez Javed Jussa Quan?ta?ve …com.estimize.public.s3.amazonaws.com/l2q presentations/3_l2q_ldn... · We inves.gate a wide range of machine learning algorithms

•  WedefineM&AIC(Informa.onCoefficient)astheSpearmanrankcorrela.onbetween:1)therankingofstocksbasedonafactor;and2)avectorofzero’sandone’swhereonesrepresentactualtakeovertargetsinthesubsequentmonth.

•  Toadjustforthe.mevaria.onofthepredic.vepower,wealsocomputetherisk-adjustedM&AICas:

5Sources:FTSERussell,S&PCapitalIQ,ThomsonReuters,WolfeResearchLuo’sQES

b)ABacktes?ngFrameworkforTakeoverPredic?on

AnExample–M&AIC,theVola?lityFactor

𝐴𝑣𝑒𝑟𝑎𝑔𝑒 𝐼𝐶/𝑆𝑡𝑎𝑛𝑑𝑎𝑟𝑑 𝐷𝑒𝑣𝑖𝑎𝑡𝑖𝑜𝑛 𝑜𝑓 𝐼𝐶 

Page 6: Luo’s QES Miguel Alvarez Javed Jussa Quan?ta?ve …com.estimize.public.s3.amazonaws.com/l2q presentations/3_l2q_ldn... · We inves.gate a wide range of machine learning algorithms

Valuefactors•  Moreexpensivebasedon

dividendyield,earningsyield,cashflowyield,tangiblebook-to-market,EBITDA/EV;but

•  Cheaperbasedonprice-to-sales,book-to-market,andrevenue/TEV

Qualityfactors•  Theyarelessprofitableonmost

metrics,withtheexcep.onofgrossprofitmargin.Aposi.vegrossmarginandanega.venetmarginmeanstargetfirmsarepar.cularlyinefficientingeneralmanagementandadministra.veac.vi.es.

•  Theyhavelowercorporategovernancestandardbutpossiblehigherdividendpayingsustainability,asreflectedbythelowerpayoutra.o.

•  Theyhaveslightlylowerfinancialleverageandlowerbankruptcyrisk

6

Sources:FTSERussell,S&PCapitalIQ,ThomsonReuters,WolfeResearchLuo’sQES

c)Tradi?onalStock-Sec?onFactors

A) Average M&A IC B) Risk Adjusted M&A IC

A) Average M&A IC B) Risk Adjusted M&A IC

Qua

lityFactors

ValueFactors

Page 7: Luo’s QES Miguel Alvarez Javed Jussa Quan?ta?ve …com.estimize.public.s3.amazonaws.com/l2q presentations/3_l2q_ldn... · We inves.gate a wide range of machine learning algorithms

Sen.mentfactors•  Takeovercompaniesaredislikedbysell-side

analystsbyallcommonmeasures.

Growthfactors•  Allthegrowthrelatedfactorsshownega.ve

M&AIC.•  Onemajorreasonthatanacquirerwantsto

buyatargetistoturnaroundaslowgrowthcompany,inordertogeneratemergersynergyandoutsizedprofit

TechnicalFactors•  Theyhavepoorpricemomentum.•  Theyhavelowerliquidity,e.g.,float

turnover,Amihudilliquidity.•  Theyaremorevola.le.•  Themostinteres.ngaspectisthattarget

companiestendtorallysharplyimmediatelybeforeM&Aannouncements(posi.veone-monthreturn),coupledbyhigherabnormaltradingvolume,andexcessKurtosis.ThisispossibleduetothefactthatM&Atransac.onsaresome.mesan.cipatedbycertainmarketpar.cipants.

7Sources:FTSERussell,S&PCapitalIQ,ThomsonReuters,WolfeResearchLuo’sQES

Tradi?onalStock-Sec?onFactors,Cont’dA) Average M&A IC B) Risk Adjusted M&A IC

A) Average M&A IC B) Risk Adjusted M&A IC

TechnicalFactors A) Average M&A IC B) Risk Adjusted M&A IC

GrowthFactors

Sen?

men

tFactors

Page 8: Luo’s QES Miguel Alvarez Javed Jussa Quan?ta?ve …com.estimize.public.s3.amazonaws.com/l2q presentations/3_l2q_ldn... · We inves.gate a wide range of machine learning algorithms

•  Form3andForm4areSECfilingsthatrelatetoinsidertrading.Everydirector,officerorownerofmorethan10%ofaclassofequitysecurityregisteredmustfilewiththeSECastatementofownershipregardingsuchasecurity.Theini.alfilingisonForm3andchangesarereportedonForm4.

•  Forunderperformingcompaniesbutexpec.ngtobeacquired,insidersmayengagemoreac.vetransac.onoftheirownstocks.

•  Thereareheavierthannormalinsidertransac.onsfromtwoyearspriortotheannouncementdateun.lsixmonthsrightbeforethedealsareannounced.

•  Thenumberofinsidertradesisactuallymuchlighteronequarterbeforetheannouncementdate,possiblybecauseeitherlock-outperiodpreventsinsidersfromtradingrightbeforetransac.ons,orinsiderswanttoavoidbeingperceivedastradingonprivateinforma.on.

8Sources:FTSERussell,S&PCapitalIQ,ThomsonReuters,EDGAR,WolfeResearchLuo’sQES

d)EDGARFilingSignals

NumberofFilingsforForm4andForm3

AverageM&AICforFilingMonthsPrior

A) Form 4 B) Form 3

0

500

1000

1500

2000

2500

0

100

200

300

400

500

600

A) Form 4 B) Form 3

Page 9: Luo’s QES Miguel Alvarez Javed Jussa Quan?ta?ve …com.estimize.public.s3.amazonaws.com/l2q presentations/3_l2q_ldn... · We inves.gate a wide range of machine learning algorithms

•  WeconstructtwosignalsbasedonthenumberofForm3andForm4filingsinthepasttwoandfiveyearlookbackwindows,excludingthemostrecentquarter.

•  Wealsoapplyanexponen.allydecaywithahalf-lifeofoneyearand2.5years,respec.vely.

•  Inaddi.onal,wedefinetwobinarysignals,whichisoneifthereisafilinginthepasttwo(orfive)yearlookbackwindow,orzerootherwise.

•  Signalsbasedonthelongerfive-yearwindowgenerallyhavebeperperformancethantheonesusingashortertwo-yearwindow.

•  Thebinarysignalsaresignificantlybeperthantheonesbasedonactualnumberoffilings.

•  WeusethenumberofForm3(andForm4)filingsintheimmediatepriormonthasanega.vefactor.

•  Themostrecentmonthsignalsaddpoten.aldiversifica.onbenefit

9Sources:FTSERussell,S&PCapitalIQ,ThomsonReuters,EDGAR,WolfeResearchLuo’sQES

Form3andForm4,Cont’d

SignalfromlastMonthAddsDiversifica?onBenefit

PerformanceoftheSignalBasedonForm3andForm4

A) Average M&A IC B) Risk Adjusted M&A IC

A) Average M&A IC B) Risk Adjusted M&A IC

Page 10: Luo’s QES Miguel Alvarez Javed Jussa Quan?ta?ve …com.estimize.public.s3.amazonaws.com/l2q presentations/3_l2q_ldn... · We inves.gate a wide range of machine learning algorithms

•  Form8-Kisaverybroadformusedtono.fyinvestorsofspecifiedeventsthatmaybeimportanttoshareholdersortheSEC.ThisisoneofthemostcommontypesofformsfiledwiththeSEC.Thereportno.fiesthepublicofeventsreportedincludingacquisi.on,bankruptcy,resigna.onofdirectorsandothermaterialevents.Form8-K/Aistheamendmentof8-Kfiling.

•  ThenumberofamendmentsinForm8-K/Aismuchlessfrequent

•  AmorefrequentfilingofForm8-K(andtoalesserextent,Form8-K/A)inthepasttwoyearsisassociatedwithahigherprobabilityofbeingacquired.

•  Alongerfive-yearlookbackwindowproducesslightlybepersignals

•  SimilartowhatweseeforForm3/Form4,forthemorepopularForm8-K,binarysignalsperformbeperthanfrequency-basedfactors.

10Sources:FTSERussell,S&PCapitalIQ,ThomsonReuters,EDGAR,WolfeResearchLuo’sQES

Form8-KandForm8-K/A

A) Form 8-K B) Form 8-K/A

A) Form 8-K B) Form 8-K/A

A) Average M&A IC B) Risk Adjusted M&A IC

Performan

ce

M&AIC,#ofM

onths

Prior

#ofFilings,Form8-Kand

Form

8-K/A

Page 11: Luo’s QES Miguel Alvarez Javed Jussa Quan?ta?ve …com.estimize.public.s3.amazonaws.com/l2q presentations/3_l2q_ldn... · We inves.gate a wide range of machine learning algorithms

•  TheSchedule(SC)13Disaformthatmustbefiledwhenapersonorgroupacquiresmorethan5%ofanyclassofacompany'sshares.Thisinforma.onmustbedisclosedwithin10daysofthetransac.on.

•  SC13GissimilartoSC13Dusedtoreportaparty'sownershipofstockthatisover5%ofthecompany.SC13Gisshorterandrequireslessinforma.onfromthefilingparty.TobeabletofileSC13GinsteadofSC13D,thepartymustownbetween5and20%inthecompany.Thepartyacquiringthestakeinthecompanymustonlybeapassiveinvestoranddoesnotintendtoexertcontrol.

11Sources:FTSERussell,S&PCapitalIQ,ThomsonReuters,EDGAR,WolfeResearchLuo’sQES

SC13DandSC13G

NumberofFilingsforSC13D,SC13D/A,SC13GandSC13G/AA) Form SC 13D B) Form SC 13D/A

C) Form SC 13G B) Form SC 13G/A

RiskAdjustedM&AIC–SC13D,SC13D/A,SC13G,SC13G/A

AverageM&AICforDifferentLagsA) Form SC 13D B) Form SC 13D/A

C) Form SC 13G B) Form SC 13G/A

Page 12: Luo’s QES Miguel Alvarez Javed Jussa Quan?ta?ve …com.estimize.public.s3.amazonaws.com/l2q presentations/3_l2q_ldn... · We inves.gate a wide range of machine learning algorithms

•  FormDEFA14AisfiledwiththeSEC,whenaddi.onaldefini.veproxymaterialsaregiventoshareholders.

•  DFAN14Acoversaddi.onaldefini.veproxysolici.ngmaterialsfiledbynon-management.

•  FormPREC14Aisfiledwhenapreliminaryproxystatementisgiventoshareholders.

•  FormPRER14Aisfiledwhenarevisedpreliminaryproxystatementisgiventoshareholders.

•  Thesefilingsarelessfrequent,strongseasonalpaperns.•  Signalfromproxyvo.ngformsseemstodecayfaster.

12Sources:FTSERussell,S&PCapitalIQ,ThomsonReuters,EDGAR,WolfeResearchLuo’sQES

ProxyFilings

NumberofFilingsforDEFA14A,DFAN14A,PRER14A,PREC14A

RiskAdjustedM&AIC,ProxyVo?ngBasedSignals

AverageM&AICforDifferentLagsA) Form DEFA14A B) Form DFAN14A

C) Form PRER14A D) Form PREC14A

A) Form DEFA14A B) Form DFAN14A

C) Form PRER14A D) Form PREC14A

0.0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

FormDEFA14A

FormDEFA14ABinary

FormPRER14A

FormPRER14ABinary

FormDFAN14A

FormDFAN14ABinary

FormPREC14A

FormPREC14ABinary

Twoyearwindow Fiveyearwindow

Page 13: Luo’s QES Miguel Alvarez Javed Jussa Quan?ta?ve …com.estimize.public.s3.amazonaws.com/l2q presentations/3_l2q_ldn... · We inves.gate a wide range of machine learning algorithms

•  Form11-Kistheannualreportoftheemployeestockpurchase,savingsandsimilarplans.

•  FormS-8istheini.alregistra.onstatementforsecuri.estobeofferedtotheemployeebenefitplans.

•  FormCTORDERstandsforconfiden.altreatmentorder.Itisaformfilledoutinaccordancetocompany’s8-K,10-Qor10-K.Thisformallowsacompanytokeepconfiden.alinforma.onasasecretforacertainperiodof.me;otherwisesuchinforma.onwouldputthefirmatadisadvantage.

•  FormCORRESPisanEDGARformtypethatregistrantscanusetosubmitnon-publicinforma.ontoSEC.

•  Form424B5mustbefiledwithintwobusinessdaysofthedetermina.ondateofanofferingpriceorthedatefirstusedfollowingtheeffec.vedateofapublicofferingorsaleofsecuri.esbythecompany.

•  FormEFFECTistheNo.fica.onsofEffec.veness.Itisapublicdeclara.onbytheSECthatapubliccompany’sregistra.onstatementhasbeenaccepted.Forsharesinthepubliccompaniestotradeontheopenmarket,theymustberegisteredbythecompanyandapprovedbySEC.

•  Mostsignalsshowastrongerpredic.vepowerwithalongerfive-yearlookbackwindow

•  Theexcep.onliesinthetwoformsrelatedtosecuri.esofferings(Form424B5andFormEFFECT).Becauseoftheshort.melineoftheseforms,aone-quarterlookbackwindowhasthebestperformance.

13Sources:FTSERussell,S&PCapitalIQ,ThomsonReuters,EDGAR,WolfeResearchLuo’sQES

OtherEDGARFilings

From424B5andFormEFFECT

AverageM&AIC

A) Average M&A IC B) Risk Adjusted M&A IC

Page 14: Luo’s QES Miguel Alvarez Javed Jussa Quan?ta?ve …com.estimize.public.s3.amazonaws.com/l2q presentations/3_l2q_ldn... · We inves.gate a wide range of machine learning algorithms

•  Foreachofthemajorfilingform,wepickonefactoramongthefourpossiblechoices(two-yearorfive-yearlookbackwindow,frequency-basedorbinarysignal).

•  ForForm3andform4,wechoosetwomoreshort-termsignalsbasedonthenumberoffilingsinthepreviousmonth.

14Sources:FTSERussell,S&PCapitalIQ,ThomsonReuters,EDGAR,WolfeResearchLuo’sQES

e)SummaryofEDGARSignals

PerformanceofEDGARFilingFactors

A) Average M&A IC B) Risk Adjusted M&A IC

Page 15: Luo’s QES Miguel Alvarez Javed Jussa Quan?ta?ve …com.estimize.public.s3.amazonaws.com/l2q presentations/3_l2q_ldn... · We inves.gate a wide range of machine learning algorithms

•  WeexaminethenumberofpastM&A,buyback,andotherrelatedac.vi.es.

•  Foreachofthethreeeventtypes,westudythreelookbackwindows:two-year,five-year,and10-year.

•  Longerlookbackwindowfactorsperformbeperthanshorterwindow.

•  Binarysignalshavecomparableperformanceasfrequency-basedones.

•  Thecorrela.onsamongthesignalsbasedonthethreeeventtypesarefairlylow.

•  Totakeadvantageofthediversifica.onbenefit,wefurtherconstructasimpleequallyweightedfactor.Thecombinedfactorbooststheperformanceevenfurther.

15Sources:FTSERussell,S&PCapitalIQ,ThomsonReuters,WolfeResearchLuo’sQES

M&AEventCountSignal

A) Average M&A IC B) Risk Adjusted M&A IC

A) Average M&A IC B) Risk Adjusted M&A IC

A) Average M&A IC B) Risk Adjusted M&A IC

Combine

dSignalOffe

rs

Diversifica?o

nBe

nefit

Performan

ceofB

inary

Signals

Performan

ceofM

&A

Even

tCou

ntSigna

ls

Page 16: Luo’s QES Miguel Alvarez Javed Jussa Quan?ta?ve …com.estimize.public.s3.amazonaws.com/l2q presentations/3_l2q_ldn... · We inves.gate a wide range of machine learning algorithms

•  OursectorM&Amomentumfactor,whichisdefinedas:.

•  Hotsectorsaremorelikelytostayhotforawhile.•  TheperformanceofoursectorM&Amomentumfactorisstrongand

uncorrelatedtothethreeM&Aeventcountsignals.

16Sources:FTSERussell,S&PCapitalIQ,ThomsonReuters,WolfeResearchLuo’sQES

f)SectorM&AMomentumSignal

PercentageofM&ATransac?ons,byGICSSector

PerformanceofDifferentEDGARFilingFactors

A) Percentage of M&A Transactions, by GICS Sector B) Autocorrelation

SectorM&AMomentumSignal,M&AIC

𝑆𝑒𝑐𝑡𝑜𝑟𝑀&𝐴𝑀𝑜𝑚𝑒𝑛𝑡𝑢𝑚 = . 𝜔𝑡 0# 𝑜𝑓 𝑀&𝐴 𝑇𝑟𝑎𝑛𝑠𝑎𝑐𝑡𝑖𝑜𝑛𝑠 𝑖𝑛 𝑡ℎ𝑒 𝑆𝑒𝑐𝑡𝑜𝑟

# 𝑜𝑓 𝑆𝑡𝑜𝑐𝑘𝑠 𝑖𝑛 𝑡ℎ𝑒 𝑆𝑒𝑐𝑡𝑜𝑟:

0

𝑡=−120

A) Correlation with M&A Events Signals B) Average M&A IC

Page 17: Luo’s QES Miguel Alvarez Javed Jussa Quan?ta?ve …com.estimize.public.s3.amazonaws.com/l2q presentations/3_l2q_ldn... · We inves.gate a wide range of machine learning algorithms

•  Wegroupthefactorscoveredintheprevioussec.onsintothreecategories:

–  Tradi.onalstock-selec.onfactors–  Non-tradi.onalEDGARfilingand

M&Aeventcountfactors–  SectorM&Amomentumsignal

•  Wenormalizeallfactors,matcheachtargetwitharandomlyselec.oncomparablestock,andstacktogetherallobserva.onsinthepast10yearsasourtrainingsample.

•  Thebaggingalgorithmisaddedtoallmachinelearningmodelsinthepaper.Weaveragepredic.onsfromthe10-foldbaggingformsourfinalforecast.

•  Weonlysectorneutralizealltradi.onalfactors:

–  Sectorneutraliza.onmodestlyimprovestheperformanceoftradi.onalfactors

–  Fornon-tradi.onalfactors,sectorneutraliza.onmakesnodifference.

17Sources:FTSERussell,S&PCapitalIQ,ThomsonReuters,WolfeResearchLuo’sQES

g)MachineLearningTakeovers

A) Average M&A IC B) Risk Adjusted M&A IC

A) Traditional Quant Factors B) Non-traditional Factors

A) Traditional Quant Factors B) Non-traditional Factors

RiskAdjustedM&AICfo

rtheLogitM

odel

AverageM&AICfo

rthe

LogitM

odel

Performan

ceIm

provem

ent

from

Bagging

Page 18: Luo’s QES Miguel Alvarez Javed Jussa Quan?ta?ve …com.estimize.public.s3.amazonaws.com/l2q presentations/3_l2q_ldn... · We inves.gate a wide range of machine learning algorithms

•  Logit:modelsthelogoddsoftheeventasalinearfunc.on.

•  CART:mostcommonlyusedclassifica.onalgorithms.Prunetotreesusingcross-valida.ontoavoidoverfisng.

•  RandomForest:randomlyselectasub-sampleofdatatobuildCARTmodels,taketheaveragefrommul.pleCARTmodel’spredic.ontoderivethefinalforecast

•  SVM:findthehyperplanethatmaximizeitsdistancesfromtheclosestdatapointineachofthetwogroups.

•  ANN:linearcombina.onofinputvariablestypicallytransformedbyanonlinearfunc.on

•  AdaBoost:itadap.velybuildsasequenceofweakclassifierstoconstructthestrongclassifier.Ineachstepofthetraining,itemphasizesonthemisclassifieddatasampletherebygraduallyimprovingtheaccuracyofclassifica.on.

18Sources:FTSERussell,S&PCapitalIQ,ThomsonReuters,WolfeResearchLuo’sQES

APrimerofMachineLearningAlgorithms

RandomForestFlowchart

SVMIllustrated

A) The growth of # of papers on SSRN B) Seasonality, # of papers posted by month

Page 19: Luo’s QES Miguel Alvarez Javed Jussa Quan?ta?ve …com.estimize.public.s3.amazonaws.com/l2q presentations/3_l2q_ldn... · We inves.gate a wide range of machine learning algorithms

•  Aswemovefroma80-factormodeltoa235-factormodel,predic.vepowerimproves.The80factorsareasubsetofthe235signals.•  Wecanfurtherboostperformancebyaddingnon-tradi.onalfactors(i.e.,EDGARfilingbased,eventcountsignals,andoursectorM&A

momentumfactor).

19Sources:FTSERussell,S&PCapitalIQ,ThomsonReuters,WolfeResearchLuo’sQES

AHorseraceofMachineLearningAlgorithmsinRareEventPredic?on

IncreasingtheNumberofTradi?onalQuantFactor

AddingNon-tradi?onalFactorsfurtherBoostsPerformance

A) Average M&A IC B) Risk adjusted M&A IC

A) Average M&A IC B) Risk Adjusted M&A IC

Page 20: Luo’s QES Miguel Alvarez Javed Jussa Quan?ta?ve …com.estimize.public.s3.amazonaws.com/l2q presentations/3_l2q_ldn... · We inves.gate a wide range of machine learning algorithms

•  Thereisanicemixofbothtradi.onalandnon-tradi.onalfactorsbeenselectedbythemodel.

•  Mostfactorsshowintui.vesigns.•  Theweightsofmostfactorsarequitestableover.me.

20Sources:FTSERussell,S&PCapitalIQ,ThomsonReuters,WolfeResearchLuo’sQES

FactorsSelectedbytheModel

FactorWeights,LogitModelatMay31,2017

WeightsfortheM&AFrequencyFactor

WeightsforthePrice-to-salesFactor

Page 21: Luo’s QES Miguel Alvarez Javed Jussa Quan?ta?ve …com.estimize.public.s3.amazonaws.com/l2q presentations/3_l2q_ldn... · We inves.gate a wide range of machine learning algorithms

•  OurfinalSMAP(Systema.cMerger&Acquisi.onPredic.on)model,whichequallyweighsthepredic.onfromsevenoftheeightmodels.WeexcludetheCARTmodel,becauseourrandomforecastmodelencompassestheCARTmodelandofferssignificantlybeperperformance.

21Sources:FTSERussell,S&PCapitalIQ,ThomsonReuters,WolfeResearchLuo’sQES

h)AssessingtheModel

SMAPModelPerformsBeierinRecentYears

M&AIC,SMAPModel

A) Average M&A IC B) Risk Adjusted M&A IC

AverageM&AIC

Risk-AdjustedM&AIC

Page 22: Luo’s QES Miguel Alvarez Javed Jussa Quan?ta?ve …com.estimize.public.s3.amazonaws.com/l2q presentations/3_l2q_ldn... · We inves.gate a wide range of machine learning algorithms

22Sources:FTSERussell,S&PCapitalIQ,ThomsonReuters,WolfeResearchLuo’sQES

Out-of-sampleHitRate

%ofActualTakeoverCaptured#ofActualM&A,byQuin?le

HitRateatVariousCutoffThresholdsHitRatebyForecastedDecileintheSubsequentMonth

Page 23: Luo’s QES Miguel Alvarez Javed Jussa Quan?ta?ve …com.estimize.public.s3.amazonaws.com/l2q presentations/3_l2q_ldn... · We inves.gate a wide range of machine learning algorithms

23Sources:FTSERussell,S&PCapitalIQ,ThomsonReuters,WolfeResearchLuo’sQES

Forecas?ngAbilityatdifferentHorizon

AverageHitRateatEachDecileforDifferentHorizonSMAPModelAutocorrela?on

Cumula?veM&AICDecayM&AICDecay

Page 24: Luo’s QES Miguel Alvarez Javed Jussa Quan?ta?ve …com.estimize.public.s3.amazonaws.com/l2q presentations/3_l2q_ldn... · We inves.gate a wide range of machine learning algorithms

24Sources:FTSERussell,S&PCapitalIQ,ThomsonReuters,WolfeResearchLuo’sQES

ACaseStudy–WholeFoodsMarket

MachineLearningModelOutput,October2016

SMAPPercen?leforWFMbeforeitsM&AAnnouncement TopFactorsthatContributetotheChangeofLogitModel,April2017

TopFactorsthatContributetotheChangeofLogitModel,October2016

Page 25: Luo’s QES Miguel Alvarez Javed Jussa Quan?ta?ve …com.estimize.public.s3.amazonaws.com/l2q presentations/3_l2q_ldn... · We inves.gate a wide range of machine learning algorithms

•  Buyingpoten.alM&Atargetsbeforedealannouncementstocapturethelargeposi.vereturnsontheM&Aannouncementdate

•  SMAPisnotatypicalstock-selec.onmodel.Thecorrela.onbetweentheSMAPandourflagshipglobalstock-selec.onmodel–LEAPisactuallynega.vemostofthe.me

•  Wecan’tseemtousetheSMAPnaivelyasastockselec.ontool.

•  ForM&Atransac.ons,theupsideisallintheoutliers,i.e.,thoseactualtakeovertargetcompanies.

•  ThetopdecilebasedontheSMAPmodelclearlyoutperformsthemarketandallotherdeciles

25Sources:FTSERussell,S&PCapitalIQ,ThomsonReuters,WolfeResearchLuo’sQES

i)Prac?calApplica?on

Correla?onbetweentheSMAPandLEAP

SMAP,RankICCumula?vePerformance,byDeciles

The top 10% stocks with the highest takeover

probabilities

Page 26: Luo’s QES Miguel Alvarez Javed Jussa Quan?ta?ve …com.estimize.public.s3.amazonaws.com/l2q presentations/3_l2q_ldn... · We inves.gate a wide range of machine learning algorithms

26Sources:FTSERussell,S&PCapitalIQ,ThomsonReuters,WolfeResearchLuo’sQES

SkewnessDrivenbyM&ATarget

ReturnDistribu?onCAGR,byDecile

MedianSkewness,byDecile

A) Return Distribution for the Top Decile SMAP B) Return Distribution for the Russell 3000 Universe

PerformancefortheSMAPTopDecilewithandwithouttheM&ATarget

A) Wealth Curve B) Sharpe Ratio

The top 10% stocks with the highest takeover

probabilities

The top 10% stocks with the highest takeover probabilities

Page 27: Luo’s QES Miguel Alvarez Javed Jussa Quan?ta?ve …com.estimize.public.s3.amazonaws.com/l2q presentations/3_l2q_ldn... · We inves.gate a wide range of machine learning algorithms

•  SMAPmodel’shitrateincreasesmonotonicallyaswe.ghtenoutcutoffthreshold.

•  Aswelitupourthreshold,ourpor8oliobecomesconcentrated.Theremightbenotakeovertargetatallincertainmonths.

•  Asthepor8oliobecomesmoreconcentra.on,thepercentageofmonthswithatleastoneM&Atransac.onalsodecreasesmonotonically

•  5%istheop.malcutoffthreshold,whichgivesus150stockspermonth.

27Sources:FTSERussell,S&PCapitalIQ,ThomsonReuters,WolfeResearchLuo’sQES

TheOp?malCutoffThreshold

PerformancefortheSMAPwithDifferentCutoffThreshold

A) Hit Rate by Various Cutoff Threshold B) Percentage of Months with at least One Deal

C) Cumulative Performance D) Sharpe Ratio

Page 28: Luo’s QES Miguel Alvarez Javed Jussa Quan?ta?ve …com.estimize.public.s3.amazonaws.com/l2q presentations/3_l2q_ldn... · We inves.gate a wide range of machine learning algorithms

28Sources:FTSERussell,S&PCapitalIQ,ThomsonReuters,WolfeResearchLuo’sQES

AvoidShor?ngPoten?alTakeoverTargets

RemovingHigh-Takeover-ProbabilityStocks(basedonSMAP)BoostsPerformance

Cumula?vePerformance,ROEFactorPorDolio ImprovingthePerformanceofLEAPusingtheSMAPModel

A) Annualized Return B) Sharpe Ratio

C) Annualized Volatility D) Max Drawdown

•  Weseeacrosstheboardperformanceimprovementbyremovingthehighesttakeoverprobabilitystocksfromtheshortside.

•  Bynotshor.ngpoten.alM&Atargetsalsoreducestheriskforallcommonfactors.

•  RemovinghightakeoverprobabilitystocksfromtheshortsidealsohelpshighefficacyalphamodelssuchastheLEAP.

A) Sharpe Ratio B) Max Drawdown

Page 29: Luo’s QES Miguel Alvarez Javed Jussa Quan?ta?ve …com.estimize.public.s3.amazonaws.com/l2q presentations/3_l2q_ldn... · We inves.gate a wide range of machine learning algorithms

29

DISCLOSURE SECTION AnalystCer?fica?on:TheanalystofWolfeResearchprimarilyresponsibleforthisresearchreportwhosenameappearsfirstonthefrontpageofthisresearchreportherebycer.fiesthat(i)therecommenda.onsandopinionsexpressedinthisresearchreportaccuratelyreflecttheresearchanalysts’personalviewsaboutthesubjectsecuri.esorissuersand(ii)nopartoftheresearchanalysts’compensa.onwas,isorwillbedirectlyorindirectlyrelatedtothespecificrecommenda.onsorviewscontainedinthisreport.OtherDisclosures:WolfeResearch,LLCdoesnotassignra.ngsofBuy,HoldorSelltothestocksitcovers.Outperform,PeerPerformandUnderperformarenottherespec.veequivalentsofBuy,HoldandSellbutrepresentrela.veweigh.ngsasdefinedabove.Tosa.sfyregulatoryrequirements,OutperformhasbeendesignatedtocorrespondwithBuy,PeerPerformhasbeendesignatedtocorrespondwithHoldandUnderperformhasbeendesignatedtocorrespondwithSell.WolfeResearchSecuri.esandWolfeResearch,LLChaveadoptedtheuseofWolfeResearchasbrandnames.WolfeResearchSecuri.es,amemberofFINRA(www.finra.org)isthebroker-dealeraffiliateofWolfeResearch,LLCandisresponsibleforthecontentsofthismaterial.AnyanalystspublishingthesereportsareduallyemployedbyWolfeResearch,LLCandWolfeResearchSecuri.es.Thecontentofthisreportistobeusedsolelyforinforma.onalpurposesandshouldnotberegardedasanoffer,orasolicita.onofanoffer,tobuyorsellasecurity,financialinstrumentorservicediscussedherein.Opinionsinthiscommunica.oncons.tutethecurrentjudgmentoftheauthorasofthedateand.meofthisreportandaresubjecttochangewithoutno.ce.Informa.onhereinisbelievedtobereliablebutWolfeResearchanditsaffiliates,includingbutnotlimitedtoWolfeResearchSecuri.es,makesnorepresenta.onthatitiscompleteoraccurate.Theinforma.onprovidedinthiscommunica.onisnotdesignedtoreplacearecipient'sowndecision-makingprocessesforassessingaproposedtransac.onorinvestmentinvolvingafinancialinstrumentdiscussedherein.Recipientsareencouragedtoseekfinancialadvicefromtheirfinancialadvisorregardingtheappropriatenessofinves.nginasecurityorfinancialinstrumentreferredtointhisreportandshouldunderstandthatstatementsregardingthefutureperformanceofthefinancialinstrumentsorthesecuri.esreferencedhereinmaynotberealized.Pastperformanceisnotindica.veoffutureresults.Thisreportisnotintendedfordistribu.onto,oruseby,anypersonoren.tyinanyloca.onwheresuchdistribu.onorusewouldbecontrarytoapplicablelaw,orwhichwouldsubjectWolfeResearch,LLCoranyaffiliatetoanyregistra.onrequirementwithinsuchloca.on.Foraddi.onalimportantdisclosures,pleaseseewww.wolferesearch.com\disclosures.TheviewsexpressedinWolfeResearch,LLCresearchreportswithregardstosectorsand/orspecificcompaniesmayfrom.meto.mebeinconsistentwiththeviewsimpliedbyinclusionofthosesectorsandcompaniesinotherWolfeResearch,LLCanalysts’researchreportsandmodelingscreens.WolfeResearchcommunicateswithclientsacrossavarietyofmediumsoftheclients’choosingincludingemails,voiceblastsandelectronicpublica.ontoourproprietarywebsite.Copyright©WolfeResearch,LLC2017.Allrightsreserved.Allmaterialpresentedinthisdocument,unlessspecificallyindicatedotherwise,isundercopyrighttoWolfeResearch,LLC.Noneofthematerial,noritscontent,noranycopyofit,maybealteredinanyway,ortransmipedtoordistributedtoanyotherparty,withoutthepriorexpresswripenpermissionofWolfeResearch,LLC.ThisreportislimitedforthesoleuseofclientsofWolfeResearch.Authorizedusershavereceivedanencryp.ondecoderwhichlegislatesandmonitorstheaccesstoWolfeResearch,LLCcontent.Anydistribu.onofthecontentproducedbyWolfeResearch,LLCwillviolatetheunderstandingofthetermsofourrela.onship.