Using Data Science & Predictive Models to Produce ForesightForesight: The case of the presumptuous...

38
Using Data Science & Predictive Models to Produce Foresight: The case of the presumptuous assumptions Steve Cohen Partner in4mation insights Needham, MA

Transcript of Using Data Science & Predictive Models to Produce ForesightForesight: The case of the presumptuous...

Page 1: Using Data Science & Predictive Models to Produce ForesightForesight: The case of the presumptuous assumptions Steve Cohen Partner in4mation insights ... The Business Intelligence

Using Data Science & Predictive Models to Produce Foresight:The case of the presumptuous assumptions

Steve CohenPartnerin4mation insightsNeedham, MA

Page 2: Using Data Science & Predictive Models to Produce ForesightForesight: The case of the presumptuous assumptions Steve Cohen Partner in4mation insights ... The Business Intelligence

Big Data Analytics Consumer & Market Segmentation Customer Lifetime Value & Churn Digital Attribution Market Structure Marketing Mix Modeling New Product & Service Design Pricing & Promotion Optimization Assortment Optimization Retail Site Location Marketing Ecosystem Models

2

Proprietary Hardware & Software

Hierarchical Bayesian Statistics

Top Marketing Science Advisors

Founders are Thought Leaders

© 2014 by in4mation insights, LLC

Page 3: Using Data Science & Predictive Models to Produce ForesightForesight: The case of the presumptuous assumptions Steve Cohen Partner in4mation insights ... The Business Intelligence

Why should you care what I have to say?

First to do Choice based Conjoint Analysis commercially (1983)

First academic paper for Latent Class CBCA (1995)

First integrated model for multiway segmentation based on Latent Class Models (1996)

First academic paper for Menu based Conjoint Analysis (2000)

Introduced MaxDiff scaling at ESOMAR (2000)

Best paper at Sawtooth Conference: MaxDiff (2003)

ESOMAR best paper of the year award for MaxDiff (2004)

Best paper in Marketing Research Magazine: MaxDiff (2005)

American Marketing Association Parlin Award (2011)

NextGen Marketing Research LinkedIn group: Individual Disruptive Innovator award (2012)

Marketing Research Council of NYC: MR Hall of Fame (2013)

Over two dozen papers and presentations at industry conferences on analytics & modeling

© 2014 by in4mation insights, LLC 3

Page 4: Using Data Science & Predictive Models to Produce ForesightForesight: The case of the presumptuous assumptions Steve Cohen Partner in4mation insights ... The Business Intelligence

What is Data Science?

Page 5: Using Data Science & Predictive Models to Produce ForesightForesight: The case of the presumptuous assumptions Steve Cohen Partner in4mation insights ... The Business Intelligence

What is Data Science?

© 2014 by in4mation insights, LLC 5

Page 6: Using Data Science & Predictive Models to Produce ForesightForesight: The case of the presumptuous assumptions Steve Cohen Partner in4mation insights ... The Business Intelligence

What is Data Science?

© 2014 by in4mation insights, LLC

Credit: Drew Conway

6

Page 7: Using Data Science & Predictive Models to Produce ForesightForesight: The case of the presumptuous assumptions Steve Cohen Partner in4mation insights ... The Business Intelligence

What is Data Science?

© 2014 by in4mation insights, LLC 7

Page 8: Using Data Science & Predictive Models to Produce ForesightForesight: The case of the presumptuous assumptions Steve Cohen Partner in4mation insights ... The Business Intelligence

Market Segmentation

Page 9: Using Data Science & Predictive Models to Produce ForesightForesight: The case of the presumptuous assumptions Steve Cohen Partner in4mation insights ... The Business Intelligence

Analytic steps in a typical segmentation study

Collect RatingsFactor

AnalysisCluster

AnalysisAssignment

Tool

Tandem Clustering

“Tandem clustering (i.e. factor analysis followed by cluster analysis) is an outmoded and statistically insupportable practice.”

Arabie & Hubert (1994)© 2014 by in4mation insights, LLC 9

Page 10: Using Data Science & Predictive Models to Produce ForesightForesight: The case of the presumptuous assumptions Steve Cohen Partner in4mation insights ... The Business Intelligence

Principle Components Analysis Principal Components Analysis

What are you doing?

© 2014 by in4mation insights, LLC 10

Page 11: Using Data Science & Predictive Models to Produce ForesightForesight: The case of the presumptuous assumptions Steve Cohen Partner in4mation insights ... The Business Intelligence

Discriminate Analysis Discriminant Analysis

What are you doing?

© 2014 by in4mation insights, LLC 11

Page 12: Using Data Science & Predictive Models to Produce ForesightForesight: The case of the presumptuous assumptions Steve Cohen Partner in4mation insights ... The Business Intelligence

Rating Scales

Factor Analysis

Cluster Analysis

What’s my beef with common segmentation practice? The short list.

Guiding Principle:

Segmentation is a search for differences

© 2014 by in4mation insights, LLC 12

Page 13: Using Data Science & Predictive Models to Produce ForesightForesight: The case of the presumptuous assumptions Steve Cohen Partner in4mation insights ... The Business Intelligence

Hierarchical Bayesian Modeling

Page 14: Using Data Science & Predictive Models to Produce ForesightForesight: The case of the presumptuous assumptions Steve Cohen Partner in4mation insights ... The Business Intelligence

��1 = ��1 +��11 ∗����������+��21 ∗����������_��������_����������+��31 ∗��ℎ����������+��41 ∗����������_������������

��2 = ��2 +��12 ∗����������+��22 ∗����������_��������_����������+��32 ∗��ℎ����������+��42 ∗����������_������������

What are the effects of price and in-store display on sales of supermarket product?

�����������������!= ��+��1 ∗��������������+��2 ∗ln(����������)

Lower ModelUpper Model

© 2014 by in4mation insights, LLC 14

Page 15: Using Data Science & Predictive Models to Produce ForesightForesight: The case of the presumptuous assumptions Steve Cohen Partner in4mation insights ... The Business Intelligence

What could effect sales of SKUs in a store?

Lower Model

National TV

Local TV

Radio

Outdoor

Magazines

Newspapers

Social media activity

Website & search

Upper Model

Channel

Geography

Ingredients

Location at point of sale

Store size

Store age

Store format

Company vs. franchise

Demos of trading area

Lower Model

Base Price

Discounted Price

Feature

Display

Form

Size

Coupons

Seasonality

Holidays

Weather

© 2014 by in4mation insights, LLC 15

Page 16: Using Data Science & Predictive Models to Produce ForesightForesight: The case of the presumptuous assumptions Steve Cohen Partner in4mation insights ... The Business Intelligence

Effectiveness Index per $1MM in Brand Size:

121 30 97 91 72 125 181 111 72

$28,328

$16,980

$11,939

$9,318

$6,828

$4,766 $4,751 $3,312

$1,518

Brand A Brand B Brand C Brand D Brand E Brand F Brand G Brand H Brand I

Bayesian analysis works best when there are many items, brands, stores or regions that need to be compared.

Category Average $9,722

121 30 97 91 72 125 181 111 72

TV Effectiveness:Sales/GRPs

Items can be compared to

average

Items can be indexed

against their volume

Growth opportunity

© 2014 by in4mation insights, LLC 16

Page 17: Using Data Science & Predictive Models to Produce ForesightForesight: The case of the presumptuous assumptions Steve Cohen Partner in4mation insights ... The Business Intelligence

Choice Modeling &Trade-Up

Page 18: Using Data Science & Predictive Models to Produce ForesightForesight: The case of the presumptuous assumptions Steve Cohen Partner in4mation insights ... The Business Intelligence

Discreet Choice Model Discrete Choice Model

What are you doing?

© 2014 by in4mation insights, LLC 18

Page 19: Using Data Science & Predictive Models to Produce ForesightForesight: The case of the presumptuous assumptions Steve Cohen Partner in4mation insights ... The Business Intelligence

True or False:Discrete Choice Models are the exact same thing as Choice-based Conjoint Analysis

Page 20: Using Data Science & Predictive Models to Produce ForesightForesight: The case of the presumptuous assumptions Steve Cohen Partner in4mation insights ... The Business Intelligence

A typology of choice models

How many brands or items chosen?

Only one More than oneH

ow

man

y u

nit

s o

f eac

h it

em

ch

ose

n?

Mo

re t

han

on

eO

nly

on

e

Dell100 GB Hard Drive

4 MB RAMBasic Processor17-inch Screen

MS Office90-day Warranty

Total Price: $1,170

HP200 GB Hard Drive

2 MB RAMEnhanced Processor

19-inch ScreenMS Office

90-day WarrantyTotal Price: $1,480

Sony Vaio500 GB Hard Drive

4 MB RAMBasic Processor14-inch ScreenNo MS Office

1 Year WarrantyTotal Price: $1,840

© 2014 by in4mation insights, LLC 20

Page 21: Using Data Science & Predictive Models to Produce ForesightForesight: The case of the presumptuous assumptions Steve Cohen Partner in4mation insights ... The Business Intelligence

Price elasticity is about substitutability

$13.99 $10.99

$12.99

$229 $249

$234

$179 $199

$199

© 2014 by in4mation insights, LLC 21

Page 22: Using Data Science & Predictive Models to Produce ForesightForesight: The case of the presumptuous assumptions Steve Cohen Partner in4mation insights ... The Business Intelligence

Trade-up happens when shoppers are willing to spend more.

$3.99

Coca-ColaClassic

6-pk, 12oz cans

$5.49

Coca-ColaClassic

12-pk, 12oz cans

$1.49

Coca-ColaClassic

20 oz bottle

$1.89

Coca-ColaClassic

2 liter bottle

Quality Count Size

Coca-ColaClassic

12-pk, 12oz cans

$5.49

Private LabelCola

12-pk, 12oz cans

$2.99

© 2014 by in4mation insights, LLC 22

Page 23: Using Data Science & Predictive Models to Produce ForesightForesight: The case of the presumptuous assumptions Steve Cohen Partner in4mation insights ... The Business Intelligence

Trade-Up Model assumptions

Products are not substitutes

Trade up/down is asymmetric

Consumers will purchase the most quantity that they can

Subject to their budget limit

Subject to diminishing returns

Having money left over after making the purchase is good

© 2014 by in4mation insights, LLC 23

Page 24: Using Data Science & Predictive Models to Produce ForesightForesight: The case of the presumptuous assumptions Steve Cohen Partner in4mation insights ... The Business Intelligence

Market share simulations: Trade-Up vs. HBCBCA Logit Model

© 2014 by in4mation insights, LLC

0%

2%

4%

6%

8%

10%

12%

14%

16%

18%

$299 $399 $499 $699 $899

Pre

dic

ted

Mark

et

Sh

are

Price of Brand A

BRAND C (Tradeup)BRAND C (HB Logit)

BRAND B (Tradeup)BRAND B (HB Logit) Market share is

predicted to be higher for Brand A in the Trade-up model.

BRAND A (Tradeup)

BRAND A (HB Logit)

24

Page 25: Using Data Science & Predictive Models to Produce ForesightForesight: The case of the presumptuous assumptions Steve Cohen Partner in4mation insights ... The Business Intelligence

Market share simulations: Trade-Up vs. HBCBCA Logit Model

© 2014 by in4mation insights, LLC

0%

2%

4%

6%

8%

10%

12%

14%

16%

18%

$299 $399 $499 $699 $899

Pre

dic

ted

Mark

et

Sh

are

Price of Brand A

Which price elasticity makes more sense?

BRAND A (Tradeup)

BRAND A (HB Logit)

25

Page 26: Using Data Science & Predictive Models to Produce ForesightForesight: The case of the presumptuous assumptions Steve Cohen Partner in4mation insights ... The Business Intelligence

Modeling the Marketing Ecosystem

Page 27: Using Data Science & Predictive Models to Produce ForesightForesight: The case of the presumptuous assumptions Steve Cohen Partner in4mation insights ... The Business Intelligence

27

The Business Intelligence Landscape is changing.

More holistic view of business needed

Increasing role of social & digital media

Fusing data sources into new databases

Mine existing data

Existing analytic tools assume static rather than dynamic view

Integrate consumer-based metrics into modeling and planning models

Need to accurately measure both short- and long-term marketing effects

Need reliable measurement of effects of traditional marketing vs. social/digital media

© 2014 by in4mation insights, LLC

Page 28: Using Data Science & Predictive Models to Produce ForesightForesight: The case of the presumptuous assumptions Steve Cohen Partner in4mation insights ... The Business Intelligence

Any time series can be modeled as a simple process, where next month is a function of previous months.

Offline & Online Marketing Tactics

BHMs & Social Metrics

Sales

© 2014 by in4mation insights, LLC 28

Page 29: Using Data Science & Predictive Models to Produce ForesightForesight: The case of the presumptuous assumptions Steve Cohen Partner in4mation insights ... The Business Intelligence

Some marketing tactics may have an immediate effect on sales, while others may take time to change opinions.

Sales

Imm

ed

iate

Offline & Online Marketing Tactics

BHMs & Social Metrics

© 2014 by in4mation insights, LLC 29

Page 30: Using Data Science & Predictive Models to Produce ForesightForesight: The case of the presumptuous assumptions Steve Cohen Partner in4mation insights ... The Business Intelligence

Once opinions have changed, time passesbefore the impact on sales may be seen.

Sales

Imm

ed

iate

Offline & Online Marketing Tactics

BHMs & Social Metrics

© 2014 by in4mation insights, LLC 30

Page 31: Using Data Science & Predictive Models to Produce ForesightForesight: The case of the presumptuous assumptions Steve Cohen Partner in4mation insights ... The Business Intelligence

Feedback occurs as higher sales affect consumer perceptions, leading to changes in consumer sentiment and more online ‘buzz” and activity.

Sales

Imm

ed

iate

Offline & Online Marketing Tactics

BHMs & Social Metrics

© 2014 by in4mation insights, LLC 31

Page 32: Using Data Science & Predictive Models to Produce ForesightForesight: The case of the presumptuous assumptions Steve Cohen Partner in4mation insights ... The Business Intelligence

Recent application

200 Millward Brown attributes & funnel metrics

20 marketing spend tactics

Number of channels = 11

Number of SKUs = 15

Number of time periods = 39 each SKU

© 2014 by in4mation insights, LLC 32

Page 33: Using Data Science & Predictive Models to Produce ForesightForesight: The case of the presumptuous assumptions Steve Cohen Partner in4mation insights ... The Business Intelligence
Page 34: Using Data Science & Predictive Models to Produce ForesightForesight: The case of the presumptuous assumptions Steve Cohen Partner in4mation insights ... The Business Intelligence

Hierarchical Bayesian statistics

Complex systems of linear or nonlinear equations

Often no analytic solution

Uses Monte Carlo simulation

Predict quantitative or qualitative phenomena

Incorporate sensible prior beliefs or knowledge

Different coefficient for each unit of analysis at the “lower”

level

“Upper” level = Context = “why behind the what”

© 2014 by in4mation insights, LLC 34

Page 35: Using Data Science & Predictive Models to Produce ForesightForesight: The case of the presumptuous assumptions Steve Cohen Partner in4mation insights ... The Business Intelligence

Big Data in, Big Data out

So, if Lower = 50 and Upper = 100, for 5,000 iterations

Confectionery ~ 3,000 SKUs 15,000,000 coefficients

Laundry products ~ 5,000 SKUs 25,000,000 coefficients

Auto Parts ~ 75,000 SKUs ~ 400 Million coefficients

Total coefficients = N_Units * (Lower + Lower * Upper)

at every iteration of the Bayesian estimation

© 2014 by in4mation insights, LLC 35

Page 36: Using Data Science & Predictive Models to Produce ForesightForesight: The case of the presumptuous assumptions Steve Cohen Partner in4mation insights ... The Business Intelligence

Not all SKUs and retailers are created equal.

Low HighLow High Low High

Lower Price Elasticity Higher

Fragrances Makeup Skin Care

AB

CD

Retaile

rs

© 2014 by in4mation insights, LLC 36

Page 37: Using Data Science & Predictive Models to Produce ForesightForesight: The case of the presumptuous assumptions Steve Cohen Partner in4mation insights ... The Business Intelligence

Trade-up has become a familiar part of the global consumer landscape.

Perceptually superior and higher price

© 2014 by in4mation insights, LLC 37

Page 38: Using Data Science & Predictive Models to Produce ForesightForesight: The case of the presumptuous assumptions Steve Cohen Partner in4mation insights ... The Business Intelligence

Behavioral Model

WhatMarketers Do

`What Consumers

Think & FeelWhat

Consumers Do

Brand Tracking

Advertising Testing

Market Response Models

Marketing Tactics

Brand Health &Social Metrics

Sales Performance

Analytic Framework

Proposed model

© 2014 by in4mation insights, LLC

Lack comprehensive

view

38