how to predict the future of shopping - ulrich kerzel @ papis connect
TRANSCRIPT
Dr Ulrich Kerzel
How to predict the future of shopping
2008: Founded by CERN Data Scientists
Since 2011: Award-winning retail solutions
2014: International expansion, predictive applications
LHC: 27km circumference
Photos: CERN, Blue YonderBlue Yonder History: Founded by CERN physicists
Our Journey
2014
Warburg Pincus commits $75m Investment
first go-live: Customer Targeting
first customer project using Cyclic Boosting algorithm
German Innovation Award
2008
company founded in Karlsruhe
NeuroBayes introduced
2011
first go-live: Replenishment
name change to “Blue Yonder”
2013
first go-live: Online Pricing
office in London, UK opened
2015
first go-live: Brick&Mortar Pricing
Technology Review: 50 most innovative companies
Gartner: Cool Vendor in Data Science
milestone: 41B decisions/week
2012
Retail Technology Award for best enterprise solution
Award Winning Retail Leader
The key to becoming a better company are better decisions. The key to better decisions is using your own data.
What is „Data Science“ ?Big Data Landscape in 2012
07/08/15
8
What is „Data Science“ ?
Source: M. Turck
What is Data Science ?
source: deathtothestockphoto.com
Business Development
Machine Learning
Statistics Causality
Data Exploration
Visualisation
Data Story Telling
Data Storage
DataQuality
Data Access
Programming ETL
07/08/15
Start with a vision….
Source: Flickr, by khegre, CC
What would the ideal use-case be? How would it transform thecompany into a predictive enterprise ? Think big! Think outside existingprocesses, constraints,…
Where do I want to be ? Where am I now ? What do I need to change in my organisation?
Business Development
07/08/15Source: Flickr, by vIZZual.com, CC
Start with the fundamental layout: • Scenario: what and why • Understand the processes leading to the decision
and what happens with the decision afterwards • Data available? Can they be accessed?
New data needed ?Data Quality ?
Define the “target”: What exactly needs to be optimised ?How to evaluate performance ?
Source: Flickr, by r2hox, CC
Business Development
Optimised approaches for:
► Discrete, semi-continuous and continuous variables
► Ordered and unordered classes
► Missing values(contains information as well!)
► Statistical significance ofvariables
► …
Influ
enci
ng fa
ctor
s
Influencing factorsdiscrete feature Continuous feature
Machine Learning… requires a lot of data
Machine Learning
MachineLearning
Prediction Decision &Action
average event
individual event
For example: This customer, this product on this dayin this specific location, this machine with a specific usage history, …
➢Non-Gaussian probability distributions ➢asymmetric error information (volatility) ➢Different localization and dispersion measures
What is special about the prediction of an individual density distribution?
Machine Learning
Predictive Applications Enable Ongoing Optimization
controller
Data delivery
Decisions & Actions
External factors
Predictive Application
Predictive Applications
» Decisions beat insights, any time. «
Prof. Dr. Michael Feindt
Supply Chain Pricing Marketing
Replenishment Optimization predicts demand and creates store orders that reduce out-of-stock and write-off rate at the same time.
Welt.de Wien.gv.at 3sat.de
Replenishment Optimization Up to 12% of perishable goods wasted in supermarkets. That’s about 20M tons of food
► High risk of OOS ► low stock
► Low risk of OOS ► large stock
99% Quantile
90%
80%
60%
50%
Out of Stock [%]
Writ
e –
off [
%]
1-3%
From prediction to business decision
Prob
abili
tyQuantity
Automation doubles impact
0%
2%
4%
5%
7%
Prescriptive Analytics, human decisions Automated Decisions, same stock levels
5% average out-of-stock rate
1% average out-of-stock rate
» The introduction of forecasting methods in replenishment is an important investment in the future of our stores and their processes. « Dr. Hendrik Haenecke, KT
Price Optimization measures price elasticity of demand and sets prices to increase sales, revenue and margin by 5-15%
Price and Demand
0
20
40
60
80
0 20 40 60 80
Demand
What is the ideal price?
Price and Demand
0
20
40
60
80
0 20 40 60 80
Demand Revenue
Price and Demand
0
20
40
60
80
0 20 40 60 80
Demand Revenue Cost
Price and Demand
0
20
40
60
80
0 20 40 60 80
Demand Revenue Cost Profit
The Ideal Price is Strategic
Strategy Ideal Price Worst Price
Maximize Demand 0 EUR 66 EUR
Maximize Revenue 16 EUR 66 EUR
Minimize Cost 66 EUR 0 EUR
Maximize Profit 28 EUR 0 EUR
What do mixed strategies look like?
Mixed Strategies Applied
0
12.5
25
37.5
50
0 20 40 60 80
Revenue Only Mostly Revenue Both Mostly Profit Profit Only
Mixed Strategies Applied
Strategy Ideal Price
Revenue Only 16 EUR
Mostly Revenue 19 EUR
Both 21 EUR
Mostly Profit 24 EUR
Profit Only 28 EUR
» A machine learning system such as Blue Yonder suits our dynamic business model. The solution helps us adjust early on to future developments. « Michael Sinn, Otto
Customer Targeting models uplift to distribute personalized advertising (direct mail, coupons) to maximize marketing ROI with small circulation.
Rule-based targeting
Based on your knowledge of the customer, define rules that describe a customer worthy of being targeted.
Data-driven targeting
Based on your data about customer behavior, identify customers that consistently drive revenue.
Traditionally, you do this….
… which leads to…
Rule based targetingHalf of the customers (identified by the experts) have received a catalog.
30% of these bought something evaluation period.
The other half of the customers didn‘t get a catalog.
Rule based targeting
Only 6% of these customers bought something
+ budget
+ target ?
Large effect found! +400% increase in turnover by sending a catalogue Action: Increase the marketing budget, send more catalogues
Rule based targeting
Implementing the action you find this ….
Rule based targeting
+ budget
+ target
only about 30% improvement…. what happened …? Where is the big improvement …?
Rule based targeting
Customer targeting: Causality
Correlation isn’t everything - Causality matters
Sending out a catalogue is a (conscious) business decision. The real question is: What difference does the catalogue make for the individual ?
X
Y
X
Y
X and Y are correlatedX and Y are correlated X can causally influence Y Y cannot causally influence X
Source: http://www.tylervigen.com/spurious-correlations
Correlation isn’t everything - Causality matters
Customer targeting: Causality
Historic data on
customer
Target
Action
who bought what, when, where, ….
decide to send outa catalogue (or not)
e.g. increased sales turnover, …
Customer targeting: Causality
Historic data on
customer
Target
Action
high correlation
but were are interested in this: who bought because wesent a catalogue ?
Customer targeting: Causality
Rule-based targeting
Based on your knowledge of the customer, define rules that describe a customer worthy of being targeted.
Data-driven targeting
Based on your data about customer behavior, identify customers that consistently drive revenue.
Causality-based targeting
Based on your data about customer behavior and advertising effectiveness, target customers that drive marketing ROI.
Customer targeting: Causality
Uplift Modeling
Sure Things buy in any case,
targeting or not
Sleeping Dogs stop buying when you target them
Persuadables only buy if they
get targeted
Lost Causes don’t buy anyway,
targeting or not
negative neutral positive
Direct Mail
Solution
• Customer Selection for Advertisement and Targeted Campaigns
• Selection of the most uplifted customers
• Individual selections for any campaign
Results
• ROI within one month
• Nearly zero reduction of revenue, due to better customer selection
Predictive applications enable the retailer of the future, giving more profitable growth, through automated decisions that turn strategy into concrete action.