James Taylor of Decision Management Solutions hosts Anunay Gupta of Marketelligence. Anunay discusses a practical approach to your very first predictive analytics model. Webinar recording available at https://decisionmanagement.omnovia.com/archives/42450

1. Getting your first
predictive model up
and running
James Taylor,
CEO

2. Decision Management is
A business discipline that builds on
existing enterprise applications to
put data to work
manage uncertainty
increase transparency
give the business control
©2009 Decision Management Solutions 2

3. 5 core principles of decisioning
Identify, separate and manage decisions
Use business rules to define decisions
Analytics to make decisions smarter
No answer is static
Decision-making is a process
©2009 Decision Management Solutions 3

4. Delivering Decision Management
3 stages to better operational decisions
Create a “closed
loop” between
operations and
Design and build analytics to
independent measure results and
decision processes drive improvement
to replace decision
Identify the points embedded in
decisions (usually operational systems
about customers)
that are most
important to your
operational success
©2009 Decision Management Solutions 4

5. About today’s presenters
Anunay Gupta
Co-founder and Head of Analytics, Marketelligent
10 years of experience in Consumer Banking, Risk
Management and Decision Management at American
Express and Citigroup
Work with clients to leverage their data for strategic
and tactical decisioning

6. What you should get out of this webinar
• English-version overview of what folks mean when they talk about business
intelligence, analytics, predictive analytics, models, scorecards, etc…
• Some idea of the mathematics and the sophisticated techniques that work
behind-the-scenes
• More importantly, real-life situations where you can leverage Predictive
Analytics to drive profitable growth in your business
• In the end, its not rocket science. However it does require specialized skills
and expertise for successful build and deployment
6

7. Agenda
• What is Predictive Analytics
• Critical Requirements for success
• Real life applications
 Direct Marketing : Maximizing ROI
 Consumer Finance : Whom to sell? What to sell? Which Channel?
 Consumer Packaged Goods : Marketing $ Optimization
• Summary
• Q and A
7

8. www.puntersgenie.com
……….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
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
Predictive Modeling
…. predict the probability of a horse winning a race
8

9. What is Predictive Analytics ?
“Use historical data to make certain predictions for the future”
Hindsight Insight Foresight
“What will happen?”
“What is happening ?” “Why is it happening ?”
“What should happen?”
 Typical MIS or BI  Business analysis  Predictive Analytics;
 Cognos; Business Objects;  behavior analysis; trends; forecasting; optimization,
Hyperion; ProClarity; etc etc etc
 Largely backward looking  Gives us insights on what  Uses past behavior to
is happening and why predict future outcomes
 Referred to by many folks
as ‘Analytics’ although it  Game changing
is not  Forward-looking
9

10. Some types of Predictive Analytics
Logistic Forecasting; Segmentation;
Regression OLS; ARIMA CHAID; CART
 Commonly used when the  Used to forecast  Used to bucket or ‘cluster’
objective is to predict a outcomes that are of a like things
binary outcome continuous nature  Each member in a cluster
 Example: will Customer X  Example: how much will is very similar to another
respond or not respond to this Customer Y spend in member in same cluster;
my marketing offer the next month? but very different from a
 Example: What is the  Example: movement of member in a different
chance Customer Y will the S&P 500 index on a cluster
dis-enroll in the next 12 weekly basis for the next  Example: Customers in a
months 12 weeks particular segment have
similar behaviors
ARIMA: Autoregressive Integrated Moving Average
CHAID: Chi-squared Automatic Interaction Detector
CART: Classification & Regression Tree
OLS: Ordinary Least Squares
10

11. Critical Requirements for Success
Business Objective
Data Expertise Culture
More data is better; Requires folks that Typically Senior
and data from are not only management buy-
varied information statisticians; but can in is critical.
sources is even also understand Successful
better business projects are top-
driven
Predictive Analytics
11

12. 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
What is attrition to me? No activity for 6
months / 2 months / etc
I want to predict which of my high tenure Customers will ‘attrite’
or ‘churn’ in the next 6 months
12

13. Analytical Framework
Business Objective:
I want to predict which of my high tenure Customers will ‘attrite’ or ‘churn’ in
the next 6 months
Past Future
-7 -6 -5 -4 -3 -2 -1 0 1 2 3 4 5 6 7
Months
1. Historical Customer transaction data Decision Period
(mob>12; transactions, interactions)
2. External data
(Credit bureaus; demographics; psychographic,
macroeconomic; etc)
Decision Point
Dec09
13

14. 1. Data Collection
Identify a suitable time period in the past to collect relevant information
Past
-25 -24 -23 -22 -21 -20 -19 -18 -17 -16 -15 -14 -13 -12 -11
Months
1. Historical Customer transaction data Decision Period
(mob>12; transactions, interactions) • Identify Attritors; label them as 1’s
2. External data • All others labeled as 0’s
(Credit bureaus; demographics; psychographic,
macroeconomic; etc)
Reference Point
July08
14

15. 2. Model Build & Deployment
Model
Raw data
Exploratory Data Variable Variable Development
& Deployment
Analysis Treatment Selection &
Sampling
Validation
 Data Preparation  Defining  Missing Value  Stepwise  OLS / Logistic /  Scorecard
 Over sampling ? dependent Treatment regression CHAID / etc development
 Reject
variable  Variable  Logit Plots  KS  Statistical paper
Inferencing  Business sense Transformation  Business Logic  Rank-ordering  Implementation
check  Variable capping code
 Multi-collinearity  Out-of-time
& Flooring Validation
 5 – 10 most
significant
variables
Ongoing Model Validation & Maintenance
15

16. 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
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
16

17. Resources & Timelines
CRISP-DM Process
20% 25%
15%
5%
25%
Business: 30%
Data: 40%
Modeling: 25%
10%
17

18. Explaining the benefits
Random w/ MIDAS Blaze™
100%
90%
• Save: 25% improvement in marketing
efficiency; leading to annual cost
% Responders Captured
80%
70% savings of $1.5MM. Same number of
Boost
60% Customers acquired
50%
Save
40% • Boost: 25% more acquired
30%
Customers with a marketing budget
20%
of $6MM.
10%
0%
0% 10% 20% 30% 40% 50% 60% 70% 80% 90% 100% • Build scenarios and optimize
% Mailbase
Sell the business impact; not the technical power !
18

19. Business Applications
• Optimize your Marketing $
Direct Marketing • Maximizing Customer Lifetime
Value
Consumer Finance • Deepen relationships by cross-sell
& up-sell
Telecom & Utilities • Retain Profitable Customers
• Risk Management & Fraud
Healthcare • Collect past-dues faster
• Predict Part Failures
Manufacturing
• Web targeting
19

20. 1. Direct Marketing
Cut marketing expenses significantly; while maintaining acquisition volumes
Random Mailing Intelligent Mailing
Response Rate: 4.5% Response Rate: 6.0%
Mailed
Mailed
Scorecard
Not Mailed
: Prospect
: Responder
Response Scorecards help in identifying Prospects/Customers to target
so as to maximize Response rates
20

21. Final Mailing Strategy
25% improvement in marketing ROI
- 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
RANDOM MAILINGS TARGETED MAILINGS
Response # Cumulative # Cumulative Marginal Cuml # Cumulative Marginal Cuml
# Prospects # Responders # Responders
Score Prospects Responders Response rate Response rate Responders Response rate 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%
Increasing 4 100,000 400,000 4,500 18,000 4.5% 4.5% 4,437 25,986 4.4% 6.5%
Response 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%
Rates 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%
21

22. Response Model Performance
10%
9% Modeled
8%
7%
Cumulative 6%
Response 5%
Rates
4% Random
3%
2%
1%
0%
1 2 3 4 5 6 7 8 9 10
Increasing
Response
Rates
If needed, marketing efficiencies can be further increased by targeting high
responding prospects
22

23. 2. Consumer Finance
What to Sell? To whom? Which Channel
Channels
Products
Customers
23

24. What is Customer Lifetime Value ?
Measuring Customer Lifetime Value
CLV 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
Known from Predict Response
existing P&L’s Rates
Acquisition
Monthly
Costs
Expenses
Customer
Net Margin Lifetime Value
Monthly Accumulated
Revenues Margin
Customer
Lifespan
Predict monthly
Spend Predict Customer
Attrition
24

25. Eg. Credit Cards
CLV(Customer1, product XY, Channel PQ) = f (P&L drivers, discount rate)
Customer / Segment
Acquisition Cost Acquisition Models:
Discount Rate -Product & Channel based
-p(Response Score)
Total Customers -p(Approval Score)
Revenue Models: Purchase Sales, $
-p(Activation)
Payment $
-p(Monthly purchase sales)
-p(Payment $) Net Credit Losses, $
-p(Attrition) Ending Loan Balances, $
Revenues
Expenses Expense Models:
Net Income (after taxes) -p(Credit Loss)
Terminal Value
Models can be built at Customer-
Discounted Net Income
level or Segment-level
Discounted Terminal Value
CLV
25

26. Eg. Credit Cards Cross-sell
Over 80MM Combinations !
4 Channels
Business
constraints
10 Products
Optimize
Target
Right Product to right
2MM Customers
Customer in the right
Channel
26

27. 3. Consumer Packaged Goods
Optimize marketing spend across channels
Marketing-Mix-Optimization
Optimize investments across Media so as to maximize Sales
Historical data is collected for sales (and/or other KPIs) and Multivariate regression analysis is used to quantify
all key Media Marketing activities incremental sales generated
$600,000 $600,000
$500,000 Past sales $500,000
performance
$400,000 $400,000
$300,000 $300,000
$200,000 $200,000
Incremental sales
$100,000 Past TV $100,000 generated by TV
activities
$0 $0
Week10
Week13
Week16
Week19
Week22
Week25
Week28
Week31
Week34
Week37
Week40
Week43
Week46
Week49
Week52
Week10
Week13
Week16
Week19
Week22
Week25
Week28
Week31
Week34
Week37
Week40
Week43
Week46
Week49
Week52
Week1
Week4
Week7
Week1
Week4
Week7
27

28. Optimally allocate Media spend to maximize Sales
Baseline Sales Magazine Incr. Sales TV Incr. Sales Daily Incr. Sales
test Magazine Spend TV Spend Dailies Spend
20 900
18 800
Media Spend, ‘000 SGD
16 700
Volume, ‘000 units
14
600
12
500
10
400
8
300
6
200
4
2 100
0 0
DEC07
DEC08
AUG07
OCT07
AUG08
OCT08
SEP07
SEP08
MAY07
MAY08
MAR07
JUN07
MAR08
JUN08
JUL07
NOV07
JUL08
NOV08
FEB08
FEB07
APR07
APR08
JAN07
JAN08
28

29. Magazine gives the highest ROI per $ spend
Incremental Sales per ‘000 SGD media spend
0.14
0.12
For every $ spend,
Magazine gives 6
0.10
times the return of
Efficiency
0.08 TV and dailies
0.06
0.04
0.02
-
Total Spends Magazine TV Daily
29

30. Key Takeaways
Predictive Analytics can be a potent weapon in
your toolbox
With increasing commoditization, it is truly the
next differentiator
It requires specialized expertise, talent
and tools to execute well
30

31. About Marketelligent
anunay.gupta@marketelligent.com www.marketelligent.com
1.201.301.2411
31

32. Decision Management Solutions
Decision Management Solutions can help you
Focus on the right decisions
Implement a blueprint
Define a strategy
For assistance, to find out more or if you have
questions
decisionmanagementsolutions.com/learnmore
©2009 Decision Management Solutions 32