Upload
wajdisfaihi7926
View
2.190
Download
0
Embed Size (px)
Citation preview
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)
www.marketelligent.com