Predictive Modeling

1
Predictive Modeling An Analytical
Approach Professor Dr.Purba Rao, Asian Institute
of Management, Philippines
2
Format of Presentation Context/examples where
predictive modeling could be applied Three
approaches to predictive modeling Heuristic
approach Statistical approach Data mining
approach Some details of the algorithms.
3
An example Think of a simple young entrepreneur,
Arjun Aryan, who lived in a little township
around Udaipur, India . Arjun inherited a
beautiful, historical, ornate, traditional
ancestral home which came with unique
Rajasthani architecture ,carved marble arches,
priceless antique collections and even a white
marble temple ,in the middle of a fountain lake,
just outside his home.
4
Arjun knew that there would be many people who
would be interested to visit such a beautiful
museum- like place, with an old world charm. So
he decided to renovate his ancestral home into a
culture tourism destination and market it as
such to the people in the surrounding region.
5
He knew he had a great product offering but he
had to set up a marketing campaign using
brochures on his culture tourism destination.
The potential target market for Arjun were urban
north Indian families , corporations operating
in the neighboring region and foreign national
tourists with their families.  
6
Arjun intended to reach these target markets by
direct marketing campaign, sending a promotional
brochure describing the facilities offered by
his enterprise. To maximize the response rate
, percentage of people actually visiting ,
Arjun tried to identify the consumer profile of
people who would be more likely to visit.So he
developed a data base of the target market
comprising people in the surrounding region.
7
This was actually the start of predictive
modeling which helps identify customer segments
with the highest future potential , the preferred
customers, who are the most profitable
customers, who are likely to purchase again
and who would spend the most .Predictive
modeling would also help to prioritize prospects
who start as potential customers and later a
certain proportion of them graduate to actual
customers.
8
PTM segments
Multiple variables
Highest probability of visit
Demographic data
Predictive modeling
Medium Probability of visit
Psychographic data
Lifestyle Data, purchase Habit data
Low Probability of visit
9
The identification and profiling of target
market segments which emerge from the
predictive modeling thus result in much more
focused and efficient marketing campaigns and
promotions. These efforts cut down on
ineffective marketing endeavors leading to
better bottom line, more effective use of
promotional budgets and thus higher profits.
10
Predictive Modeling usually comprises three
approaches (1) Heuristic approach ,such as RFM
analysis (2) Statistical approach like, Logistics
Regression,cluster analysis and Conjoint
Analysis (3) Data mining approach such as CHAID
Chi-square automatic interaction detection
such as the one given in Answer Tree algorithm.
11
RFM analysis could be carried out using
different Commonly available softwares such as
EXCEL , SYBIL, etc. However, SPSS works very
well for RFM segmentation. Logistics Regression,
Cluster Analysis, Discriminant Analysis can all
be carried out by SPSS. Data Mining approaches
like CHAID, can be done by Answer Tree Software
and by Clementine.
12
RFM analysis An example To give an example,
think of an owner of a hair salon offering
various hair care, make up and
curl-perm-relax-rebond kinds of products. Let
us say the owner now wants to know the profiles
of consumers who are of most value to his hair
salon. Such identification of most valuable
and preferred customers would help him to offer
special discount coupons to such customers to
show his appreciation, (b) to ensure loyalty and
also to (c )send flyers etc for a new product he
has been planning to start. Predictive modeling
can determine the critical factors which most
significantly affect his business.
13
The fundamental premise underlying RFM analysis
is that customers who have purchased recently ,
have made more purchases and have made larger
purchases are more likely to respond to your
offering than other customers who have purchased
less recently, less often and in smaller
amounts. Charlotte Mason, 2003, University of
North Carolina . The analysis helps an
organization to focus on a smaller section of
the target population which again follows another
managerial premise, Pareto Principle that 80
of the business comes from 20 of the customers
.
14
RFM in banking and wealth management RFM analysis
is extensively used in CRM (customer
relationship marketing ) and direct marketing
for selecting which consumer segments to target
for special offers, building long term
relationships in banking and wealth management
scenarios and so on. This analysis requires an
existing data base comprising the three
variables recency,(R) , how long ago the
customer last made a purchase frequency, (F),
how many times the customer has made a purchase
,within a specific time period and
monetary,(M), the total amount of money spent
by the customer within a specified time period.
15
Example of RFM database in a large
bookstore Determine profile of customers who
should be considered as preferred customers Data
fields Customer ID Age Income Gender Month last
purchased (R) Average no purchased/year
(F) Average amt spent/year (M)
16
PTM segments
Three variables
Highest Response rate
recency
Medium Response rate
RFM analysis
frequency
monetary
Low Response rate
17
Statistical approach by Logistics
Regression Here in addition to R,F,M variables
we use many other variables in the database,
which are called predictor variables.There is
also a dependent variable(D) of the dichotomous
type. In a real estate example, say we need to
determine profile of customers who would like
to Buy a townhouse. Looking at the database, in
the past if customer has ever purchased a
townhouse, D1, otherwise D0.
18
Probability of purchase
X annual income
19
Probability of purchase eZ
-------------- 1 eZ Zb0b1X1b2X2b
3X3. Z predicted score Dactual 0, 1
variable Based on Z-scores, we get probability
of purchase. Then we create deciles of consumer
profiles who are most likely to buy
20
PTM segments
Multiple variables
Highest probability Of purchase
Demographic data
Medium Probability Of purchase
Logistics regression
Psychographic data
Lifestyle Data, purchase Habit data
Low Probability Of purchase
21
Logistics Regression can be used very
effectively to predict the probability of churn
in a marketing scenario or even probability of
attrition in a BPO. Predicting Churn Predictive
modeling can help in identifying customers who
might churn and indicate to the company that
such customers would need special attention and
retention programs to win their interests
back.
22
Predicting attrition in contact center
industry. In this industry the logic of the
business depends on how efficiently and how well
the customer calls are handled or technical
support is provided by the customer service
representatives, CSRs or seat agents. However the
most important challenge faced by the industry
is the attrition phenomenon by which the CSRs
leave the organization and join competition in
huge volumes, sometimes at the rate of more than
80 a week
23
Predictive modeling can help in identifying
profiles of CSRs who would be more likely to
go for attrition . In fact among the existing
CSRs we can predict who are likely to go for
attrition and special retention programs may be
extended to them, provided they are of value to
the company. Such retention programs often are
able to win back employee loyalty and prevent
them from leaving.
24
Predictive Modeling in the credit card
industry Case of a famous bank in India which
has a credit card which is very popular. The bank
wanted to know the profile of credit card holders
who would tend not to pay the dues. For this
problem the bank already had a huge a data base
having different fields associated with the
profiles of credit card holders. Using the
database the bank identified the profile of
people who would default
25
Under Statistical approach there are several
other models also which apply very effectively.
These are Cluster Analysis, Linear Discriminant
analysis, Conjoint Analysis predicting
expected market share for a specific product
specification
26
Cluster analysis is like segmenting the
market and finding out which segment would have a
higher response rate. For this approach also we
consider demographic as well as
psycho-graphic fields, which are plotted on a
n-dimensional space.
27
x2
x5
cluster2
cluster1
x1
cluster3
x3
x4
28
Data Mining Approach CHAID Conceptually here
also you need a vast and Comprehensive data base
.Application of Chi-square based decision tree
approach commonly called CHAID helps to
identify profiles of priority customer who would
exhibit most responsiveness to the product you
have in mind.
29
CHAID Data Mining approach/Decision Tree
approach. CHAID is a combination of heuristic as
well as statistical method which examines
relationships between many categorical predictor
variables and a categorical, usually nominal,
target variable. It applies the Chi-square test
on independence, also called Contingency table
,between the target variable and each of the
predictor independent variable using the
multi-way cross tab table . The null hypothesis
H0 the two variables are independent .
30
For the decision tree models one takes a
different approach where the data set is
successively partitioned .based on relationships
between a nominal , usually, dichotomous, 0-1
, binary variable , often called target variable
or dependent variable and a large number of
independent variables, also called predictor
variables. Whenever, significant relationships
are found between target variable and different
independent variables , the decision tree
approach determines the most significant
relationship and identifies the predictor
variable which is most strongly related to the
target variable.
31
This iterative process works with repeated
application of Chi square test between target
variable Y and each one of the different
predictor variables. The predictor variable
which gives the smallest p-value provides the
basis for first partition from the root node.
Thereafter the tree grows following the same
iterative process of partitioning by the
Chi-square testing. The process of identifying
the predictor variable with the smallest p-value
is called the Bon Ferroni approximation.
32
Several computer softwares are available for the
CHAID segmentation. Answer Tree
/SPSS Clementine/SPSS The following Tree
diagram shows the segmentation process complete
with the root node and the terminal nodes.
33
(No Transcript)
34
Conjoint Analysis. A non traditional multivariat
e predictive modeling method Conjoint analysis
attempts to determine the relative importance
consumers attach to salient attributes and the
utilities they attach to the levels of
attributes. Malhotra,N.K.(Introduction to
Marketing Research,1996)
one of the greatest approaches to estimate the
market share for a new product being introduced
in the market
35
The utilities are derived from consumers
evaluation Of brands, or brand profiles composed
of these attributes and levels. The utilities and
relative importance weights help to (1)Which
attributes are important in influencing
Consumer choice (2) Determine composition of
most preferred brands (3) Estimate market shares
of brands which differ in attribute levels,
and (4) Estimate the market share for a new
product
36
Car example Attributes quality price image
environmental dimension ----------------------
--------------------------------------------------
-------------- Quality fuel
efficiency Speed Safety Comfort ---
--------------------------------------------------
---------------------------------- Price high
Medium Low -----------------------------
--------------------------------------------------
----------- Image elegance Convenience
Sporty ----------------------------------------
--------------------------------------------------
- Environmental dimension
recyclability of parts Reduced
particulate CNG option
37
Conjoint analysis creates a subset of car
profiles ( cards) by a method called Fractional
factorial design. This subset is called an
orthogonal array. These cards are ranked by
respondents in the order of preference. From
these rankings marketer gets the
relative importance of the different attributes
38
(No Transcript)
39
(No Transcript)
40
Knowing the utilities of each attribute and each
level the marketer is able to predict the
expected preference of each product profile
and ultimately the expected market share. Thus
if an entrepreneur is contemplating a new
product profile using conjoint can predict the
expected consumer preference/acceptance of the
product and the expected market share.