A primer on Predictive Analytics

Agenda

• What is Predictive Analytics

• Critical Requirements for success

• Real life applications

2

• Real life applications

� Direct Marketing : Maximizing ROI

� Consumer Finance : Whom to sell? What to sell? Which Channel?

� Consumer Packaged Goods : Marketing $ Optimization

• Summary

……….we take as much historical data from racing as we can and try to find the

things that are important for predicting the outcome of future races. Once

we find those things (in some cases we can be working with tens of

thousands of combinations of variables), we then run the models against a

test set of races and look at the results. We then look at the races that we

predicted correctly and work out what things made that possible for those

www.puntersgenie.com

Predictive Modeling

…. predict the probability of a horse winning a race

3

predicted correctly and work out what things made that possible for those

particular races. This is how we come up with the Bet Index. This

information is then fed back into the models to make them better

What is Predictive Analytics ?

“Use historical data to make certain predictions for the future”

Hindsight“What is happening ?”

Hindsight“What is happening ?”

Insight“Why is it happening ?”

Insight“Why is it happening ?”

Foresight“What will happen?”

Foresight“What will happen?”

4

“What is happening ?”“What is happening ?” “Why is it happening ?”“Why is it happening ?”“What will happen?”

“What should happen?”

“What will happen?”

“What should happen?”

• Typical MIS or BI

• Cognos; Business Objects; Hyperion; ProClarity; etc

• Largely backward looking

• Referred to by many folks as ‘Analytics’ although it is not

• Business analysis

• behavior analysis; trends; etc

• Gives us insights on what is happening and why

• Predictive Analytics; forecasting; optimization, etc

• Uses past behavior to predict future outcomes

• Game changing

• Forward-looking

• Commonly used when the objective is to predict a binary outcome

• Used to forecast outcomes that are of a continuous nature

• Used to bucket or ‘cluster’ like things

• Each member in a cluster

Some types of Predictive Analytics

Logistic

Regression

Logistic

Regression

Forecasting;

OLS; ARIMA

Forecasting;

OLS; ARIMA

Segmentation;

CHAID; CART

Segmentation;

CHAID; CART

binary outcome

• Example: will Customer X respond or not respond to my marketing offer

• Example: What is the chance Customer Y will dis-enroll in the next 12 months

continuous nature

• Example: how much will this Customer Y spend in the next month?

• Example: movement of the S&P 500 index on a weekly basis for the next 12 weeks

• Each member in a cluster is very similar to another member in same cluster; but very different from a member in a different cluster

• Example: Customers in a particular segment have similar behaviors

5

ARIMA: Autoregressive Integrated Moving Average

CHAID: Chi-squared Automatic Interaction Detector

CART: Classification & Regression Tree

OLS: Ordinary Least Squares

ARIMA: Autoregressive Integrated Moving Average

CHAID: Chi-squared Automatic Interaction Detector

CART: Classification & Regression Tree

OLS: Ordinary Least Squares

Critical Requirements for Success

Business ObjectiveBusiness Objective

6

Predictive AnalyticsPredictive Analytics

Data

More data is better;

and data from

varied information

sources is even

better

Data

More data is better;

and data from

varied information

sources is even

better

Expertise

Requires folks that

are not only

statisticians; but can

also understand

business

Expertise

Requires folks that

are not only

statisticians; but can

also understand

business

Culture

Typically Senior

management buy-

in is critical.

Successful

projects are top-

driven

Culture

Typically Senior

management buy-

in is critical.

Successful

projects are top-

driven

Business Objective

I want to identify which Customers will ‘attrite’ so that I can take some

proactive actions

All Customers? Or just new Customers???

Attrite today / tomorrow / next month / etc

7

I want to predict which of my high tenure Customers will ‘attrite’

or ‘churn’ in the next 6 months

Attrite today / tomorrow / next month / etc

What is attrition to me? No activity for 6

months / 2 months / etc

Analytical Framework

Business Objective:I want to predict which of my high tenure Customers will ‘attrite’ or ‘churn’ in

the next 6 months

FuturePast

8

Decision Period

Months

Decision PointDec09

-7 -6 -5 -4 -3 -2 -1 0 1 2 3 4 5 6 7

1. Historical Customer transaction data(mob>12; transactions, interactions)

2. External data(Credit bureaus; demographics; psychographic,

macroeconomic; etc)

1. Data Collection

Past

Identify a suitable time period in the past to collect relevant information

9

Decision Period

• Identify Attritors; label them as 1’s

• All others labeled as 0’s

Months

Reference PointJuly08

2. External data(Credit bureaus; demographics; psychographic,

macroeconomic; etc)

-25 -24 -23 -22 -21 -20 -19 -18 -17 -16 -15 -14 -13 -12 -11

1. Historical Customer transaction data(mob>12; transactions, interactions)

2. Model Build & Deployment

• Data Preparation

• Over sampling ?

Raw data

&

Sampling

Raw data

&

Sampling

• Defining dependent variable

Exploratory Data

Analysis

Exploratory Data

Analysis

• Missing Value Treatment

Variable

Treatment

Variable

Treatment

• Stepwise regression

Variable

Selection

Variable

Selection

• OLS / Logistic / CHAID / etc

Model

Development

&

Validation

Model

Development

&

Validation

• Scorecard development

DeploymentDeployment

10

• Over sampling ?

• Reject Inferencing

variable

• Business sense check

• Variable Transformation

• Variable capping & Flooring

• Logit Plots

• Business Logic

• Multi-collinearity

• 5 – 10 most significant variables

• KS

• Rank-ordering

• Out-of-time Validation

• Statistical paper

• Implementation code

Ongoing Model Validation & Maintenance

Output of Modeling Process

Every Customer has a unique ‘Score’ that captures the essence of

what is being modeled.

The ‘Score’ is essentially the ‘probability’ of something happening scaled in a

pre-defined fashion; having an upper- and an lower-bound

11

pre-defined fashion; having an upper- and an lower-bound

Called a ‘Score-card’

For Example:

1. Customer #17523 has a score of 769; translating to a 90% probability of ‘churning’ in the next 6

months

2. Household # 845 has a score of 423; translating to a 36% chance of accepting the offer for a

magazine if sent a Direct mail Offer

Resources & Timelines

20% 25%

15%CR

ISP

-DM

Pro

cess

12

15%

25%

10%

5%

CR

ISP

Business: 30%

Data: 40%

Modeling: 25%

Business: 30%

Data: 40%

Modeling: 25%

Explaining the benefits

50%

60%

70%

80%

90%

100%

Random w/ MIDAS Blaze™

% R

esp

on

ders

Ca

ptu

red

• Save: 25% improvement in marketing

efficiency; leading to annual cost

savings of $1.5MM. Same number of

Customers acquiredBoost

13

0%

10%

20%

30%

40%

50%

% R

esp

on

ders

Ca

ptu

red

0% 10% 20% 30% 40% 50% 60% 70% 80% 90% 100%

% Mailbase

• Boost: 25% more acquired

Customers with a marketing budget

of $6MM.

• Build scenarios and optimize

Sell the business impact; not the technical power !

Save

• Optimize your Marketing $

• Maximizing Customer Lifetime

Value

• Deepen relationships by cross-sell

& up-sell

Direct MarketingDirect Marketing

Consumer FinanceConsumer Finance

Business Applications

14

& up-sell

• Retain Profitable Customers

• Risk Management & Fraud

• Collect past-dues faster

• Predict Part Failures

• Web targeting

Telecom & UtilitiesTelecom & Utilities

HealthcareHealthcare

ManufacturingManufacturing

Random Mailing

Response Rate: 4.5%

Ma

ile

d

Ma

ile

d

1. Direct MarketingCut marketing expenses significantly; while maintaining acquisition volumes

Intelligent Mailing

Response Rate: 6.0%

15

Response Scorecards help in identifying Prospects/Customers to target

so as to maximize Response rates

No

t M

ail

ed

Ma

ile

d

Scorecard

: Responder: Prospect

- 6 campaigns of 1MM mailings each; annual cost of $6MM

- Random mailing Response rate of 4.5% → 270,000 Responders

- Response Model built; assigns each prospect a ‘Response Score’, between 1 and 10

- 9 campaigns of 0.5MM mailings each; annual cost of $4.5MM → 270,000 Responders

- 25% improvement in marketing efficiency; leading to annual cost savings of $1.5MM

Final Mailing Strategy25% improvement in marketing ROI

16

- 25% improvement in marketing efficiency; leading to annual cost savings of $1.5MM

Increasing

Response

Rates

Response

Score# Prospects

# Cumulative

Prospects# Responders

# Cumulative

Responders

Marginal

Response rate

Cuml

Response rate# Responders

# Cumulative

Responders

Marginal

Response rate

Cuml

Response rate

1 100,000 100,000 4,500 4,500 4.5% 4.5% 9,507 9,507 9.5% 9.5%

2 100,000 200,000 4,500 9,000 4.5% 4.5% 6,761 16,268 6.8% 8.1%

3 100,000 300,000 4,500 13,500 4.5% 4.5% 5,282 21,549 5.3% 7.2%

4 100,000 400,000 4,500 18,000 4.5% 4.5% 4,437 25,986 4.4% 6.5%

5 100,000 500,000 4,500 22,500 4.5% 4.5% 4,014 30,000 4.0% 6.0%

6 100,000 600,000 4,500 27,000 4.5% 4.5% 3,592 33,592 3.6% 5.6%

7 100,000 700,000 4,500 31,500 4.5% 4.5% 3,169 36,761 3.2% 5.3%

8 100,000 800,000 4,500 36,000 4.5% 4.5% 2,958 39,718 3.0% 5.0%

9 100,000 900,000 4,500 40,500 4.5% 4.5% 2,746 42,465 2.7% 4.7%

10 100,000 1,000,000 4,500 45,000 4.5% 4.5% 2,535 45,000 2.5% 4.5%

1,000,000 45,000 4.5% 45,000 4.5%

RANDOM MAILINGS TARGETED MAILINGS

4%

5%

6%

7%

8%

9%

10%

Cumulative

Response

RatesRandom

Modeled

Response Model Performance

17

0%

1%

2%

3%

4%

1 2 3 4 5 6 7 8 9 10

Random

Increasing

Response

Rates

If needed, marketing efficiencies can be further increased by targeting high

responding prospects

2. Consumer FinanceWhat to Sell? To whom? Which Channel

Channels

Products

18

Customers

What is Customer Lifetime Value ?

Measuring Customer Lifetime ValueCLV is defined as the sum of cumulated Cash-flows – discounted using the Weighted Average

Cost of Capital (WACC) – of a Customer over his or her entire lifetime with the Franchise

Predict Response

Rates

Known from

existing P&L’s

19

Monthly

Expenses

Monthly

Expenses

Monthly

Revenues

Monthly

Revenues

Customer

Lifespan

Customer

Lifespan

Net MarginNet Margin

Accumulated

Margin

Accumulated

Margin

Acquisition

Costs

Acquisition

Costs

Customer

Lifetime Value

Customer

Lifetime Value

Predict monthly

Spend Predict Customer

Attrition

CLV(Customer1, product XY, Channel PQ) = f (P&L drivers, discount rate)CLV(Customer1, product XY, Channel PQ) = f (P&L drivers, discount rate)

Customer / Segment

Acquisition Cost

Discount Rate

Total Customers

Purchase Sales, $

Acquisition Models:-Product & Channel based

-p(Response Score)

-p(Approval Score)

Acquisition Models:-Product & Channel based

-p(Response Score)

-p(Approval Score)Revenue Models:-p(Activation)

Revenue Models:-p(Activation)

Eg. Credit Cards

20

Payment $

Net Credit Losses, $

Ending Loan Balances, $

Revenues

Expenses

Net Income (after taxes)

Terminal Value

Discounted Net Income

Discounted Terminal Value

CLV

-p(Activation)

-p(Monthly purchase sales)

-p(Payment $)

-p(Attrition)

-p(Activation)

-p(Monthly purchase sales)

-p(Payment $)

-p(Attrition)

Expense Models:-p(Credit Loss)

Expense Models:-p(Credit Loss)

Models can be built at Customer-

level or Segment-level

Eg. Credit Cards Cross-sell

4 Channels

10 Products

Over 80MM Combinations !

Optimize

Business

constraints

21

2MM Customers

Optimize

Right Product to right

Customer in the right

Channel

Target

3. Consumer Packaged GoodsOptimize marketing spend across channels

$600,000 $600,000

Historical data is collected for sales (and/or other KPIs) and

all key Media Marketing activitiesMultivariate regression analysis is used to quantify

incremental sales generated

Marketing-Mix-OptimizationOptimize investments across Media so as to maximize Sales

22

$0

$100,000

$200,000

$300,000

$400,000

$500,000

$600,000

We

ek

1

We

ek

4

We

ek

7

We

ek1

0

We

ek1

3

We

ek1

6

We

ek1

9

We

ek2

2

We

ek2

5

We

ek2

8

We

ek3

1

We

ek3

4

We

ek3

7

We

ek4

0

We

ek4

3

We

ek4

6

We

ek4

9

We

ek5

2

$0

$100,000

$200,000

$300,000

$400,000

$500,000

$600,000

We

ek

1

We

ek

4

We

ek

7

We

ek

10

We

ek

13

We

ek

16

We

ek

19

We

ek

22

We

ek

25

We

ek

28

We

ek

31

We

ek

34

We

ek

37

We

ek

40

We

ek

43

We

ek

46

We

ek

49

We

ek

52

Past TV

activities

Past sales

performance

Incremental sales

generated by TV

700

800

900

14

16

18

20

Baseline Sales Magazine Incr. Sales TV Incr. Sales Daily Incr. Sales

test Magazine Spend TV Spend Dailies Spend

Vo

lum

e,

‘00

0 u

nit

s

Me

dia

Sp

en

d,

‘00

0 S

GD

Optimally allocate Media spend to maximize Sales

23

0

100

200

300

400

500

600

0

2

4

6

8

10

12

14

JAN

07

FEB

07

MA

R0

7

AP

R0

7

MA

Y0

7

JUN

07

JUL0

7

AU

G0

7

SEP

07

OC

T0

7

NO

V0

7

DE

C0

7

JAN

08

FEB

08

MA

R0

8

AP

R0

8

MA

Y0

8

JUN

08

JUL0

8

AU

G0

8

SEP

08

OC

T0

8

NO

V0

8

DE

C0

8

Vo

lum

e,

‘00

0 u

nit

s

Me

dia

Sp

en

d,

‘00

0 S

GD

0.10

0.12

0.14

Effi

cie

ncy

Incremental Sales per ‘000 SGD media spend

Magazine gives the highest ROI per $ spend

For every $ spend,

Magazine gives 6

times the return of

TV and dailies

24

-

0.02

0.04

0.06

0.08

Total Spends Magazine TV Daily

Effi

cie

ncy

TV and dailies

Key Takeaways

Predictive Analytics can be a potent weapon in

your toolbox

With increasing commoditization, it is truly the

25

With increasing commoditization, it is truly the

next differentiator

It requires specialized expertise, talent

and tools to execute well

www.marketelligent.com

26

Thank You

27

Contact us at:

+91-80-26642802 (India)

1-201-301-2411 (USA)

info@marketelligent.com

www.marketelligent.com