Financial Services the way we see it

Predictive Modeling Using

Transactional Data
 Contents

1 Introduction 3

2 Using Transactional Data 4

3 Data Quality 4
3.1 Data Profiling 4

3.2 Exploratory Data Analysis 6

4 Cohort and Trend Analysis 7

5 Model Variable Definition 9

6 Model Selection 10

7 Conclusion 11

2
 the way we see it

1 Introduction
In a world where traditional bases of competitive advantages have dissipated,
The real benefit of analytics driven processes may be one of the few remaining points of differentiation
analytics is in using past for firms in any industry1. This is particularly true in financial services, which has
progressed rather fast along the analytical path in the last couple of decades.
data to forecast or predict
future events, providing Analytics can be used to slice and dice historical data to analyze past performance
firms with a strategic and to produce reports. Here analytics helps firms react to past events. The
real benefit of analytics is in using past data to forecast or predict future events,
capability to be proactive. providing firms with a strategic capability to be proactive.

Figure 1: Reactive vs. Proactive Decision Making

VALUE
Prediction
Forecasting
KNOWLEDGE Models

Monitoring
ANALYTICS Dashboards
Scorecards

INFORMATION
Analysis OLAP
Visualization
CONTEXT

DATA Reports Query

Search

Source: Capgemini

Predictive modeling involves creating a model that outputs the probability of an

outcome given current state values of input parameters. In banking and insurance
industries, it is typically used in the context of predicting customer behavior. Historical
data related to past customer activity is used to create a predictive model that captures
attributes which seem to have greatest influence on future customer activity.

This provides marketing departments with a great tool to optimize their marketing
campaigns, channel performance, customer on-boarding and cross-sell. These
are typically driven by predictive models for customer life-time value, behavioral
segmentation and attrition.

Figure 2: Customer Strategy driven by Predictive Analytics

Customer Lifetime Value (LTV) Behavioral Segmentation Attrition

Product
■ Estimate of customers future potential ■ The predictive models provide a behavior based The customer attrition model will provide
■
Propensity
revenue based on historical behaviors, segmentation strategy that predicts which customers the FI with an understanding of which
Index product purchase propensity and are most likely to need which products or increase customers are most likely to attrite within
credit bureau behaviors usage of current products now and in the near future the next six months

On-boarding Enterprise Cross-sell

Customer
■ The On-boarding strategy is driven by the ■ Enterprise cross-sell is driven by attrition risk, behavioral segmentation output, LTV and
Relationship
LTV, behavioral segmentation’s predictions price and channel optimization
Strategy and events based triggers
■ The strategy includes price and channel preference behaviors

Source: Capgemini

1
Competing on Analytics: The New Science of Winning by Thomas H. Davenport, Jeanne G. Harris. Harvard Business School Press
Predictive Modeling Using Transactional Data 3
 2 Using Transactional Data
A customer’s historical activity typically comprises of a few accounts and
transactions around those accounts. For example, a customer may have a checking
and savings account, a mortgage loan and a credit card from a bank. Banks also
offer services like Electronic Bill Pay (EBP) and ATM/debit cards which generate
Electronic Funds Transfer (EFT) transactions.
Transactional data
Data associated with accounts are typically stored in an Accounts Processing
potentially offers (AP) system. They may contain transactions, but AP systems usually carry only
additional levels of the last month’s history. Prior months’ transactions are reflected in monthly
insight into customer’s balance snapshots.
activity, but poses some Unlike AP data, transaction data is typically maintained as is in corresponding
challenges that need transaction processing systems, whether it is EBP or EFT. Banks may have
to be addressed before many months or years worth of daily transactional data archived and stored.
Therefore, transactional data potentially offers additional levels of insight into
analytics can derive customer’s activity.
valuable insights from it.
The richness of transactional data poses some challenges that need to be addressed
before analytics can derive valuable insights from it. The rest of this paper
details these challenges and possible solutions by referring to a case study as an
illustrative example.

3 Data Quality
As with any kind of data for any kind of analytics, data quality is the first issue to
be tackled. In order to understand the structure of data and identify issues, the key
steps are to perform data profiling and exploratory data analysis.

3.1. Data Profiling

Data profiling involves creating summary statistics for each and every column and
looking at simple plots of the data to identify trends, clusters or outliers. Summary
statistics can include count, number of missing records, mean / mode / median
values, ranges and quartiles. Box plots are useful tools to visualize some of this
information graphically.

Data profiling helps understand which columns warrant additional attention from
data quality perspective. The appropriate course of action for each column has
to be carefully determined. For some columns, missing values may be replaced by
mean or mode or a constant. Some columns may need to be simply dropped
from analysis.

4
 the way we see it

Figure 3: Box Plots to identify clusters and outliers

TRANCNT ATTRITIONIND MONTHINDEX FNCLACCTTYPEIDDCNT FNCLACCTTYPENMDCNT
-0.5 10 2 2
60

Values

Values

Values

Values

Values
40
0 5 1 1
20
0 -0.5 0 0 0
0 1 0 1 0 1 0 1 0 1
BLRORGINDVIDDCNT BLRTYPEIDDCNT BLRTIERRNKDCNT CRCARDTYPEIDDCNT CRCARDTYPEABBRVTNMDCNT
1 1 1 1 1
Values

Values

Values

Values

Values
0.5 0.5 0.5 0.5 0.5

0 0 0 0 0
0 1 0 1 0 1 0 1 0 1

PYEEORGINDVIDDCNT SPSRORGINDVID PYMTSTATIDDCNT MEDIATYPEIDDCNT PYMTTYPEDIDDCNT

20 1 2 8 6
6
4
Values

Values

Values

Values

Values
10 0.5 1 4
2
2
0 0 0 0 0
0 1 0 1 0 1 0 1 0 1
PYMTDLRAMTSUM PYMTDLRAMTAVG RECURPYMTFLGDCNT FNDGFNCLACTTYPEIDDCNT EBILLIDDCNT
10 2 2
6000 10
Values

Values

Values

Values

Values
5 4000 1 1
5
2000
0 0 0 0 0
0 1 0 1 0 1 0 1 0 1
PYMTRQSTNBRDCNT RISKOWNIDDCNT MEDIACTGYIDDCNT LGCYPOSTCDDCNT FIRSTPYEESETDTMTHS
0.5 0.5 4 1 60
40
Values

Values

Values

Values

Values
0 0 2 0.5
20

-0.5 -0.5 0 0 0
0 1 0 1 0 1 0 1 0 1

Source: Capgemini

The next step is to look further into the columns at the values represented by
the data and identify any inconsistency. For example, in a transaction file, the
transaction date cannot be earlier than the customer’s account start date. There
may also be subtle issues that cannot be caught by such logic, but can be observed
simply by plotting the corresponding attribute. As an example, the plot below
shows the number of customers who attrited each month from a bank.

In this case, the spike was caused by default values entered for some customers
whose data was migrated from one source system to another. The resolution in this
case was to not rely on the end date provided in the data column, but to define
attrition as a period of inactivity as depicted by the transaction data.

This definition also opens up the possibility of defining and detecting lower levels of
customer engagement that typically precedes attrition. Instead of defining attrition
as period of no activity, it could be defined as a period of declining activity.

Figure 4: Data Quality issue identified using a trend plot

Attrition Rate Data Quality

0.09
0.08
0.07
0.06
0.05
0.04
0.03
0.02
0.01
0
200801

200802

200803

200804

200805

200806

200807

200808

200809

200810

200811

200812

200901

200902

200903

200904

200905

200906

200907

200908

200909

200910

200911

Source: Capgemini

Predictive Modeling Using Transactional Data 5

 3.2. Exploratory Data Analysis
In exploratory data analysis, data is examined further to identify attributes that
seem significant or anomalous. This step also involves creating derived attributes
by applying transformations to original data columns. The simplest of such
transformations would be computing an Age attribute from a Birth-Date column by
differencing against current date.

For transactional data, this step often implies rolling up daily transactions into a
weekly or monthly aggregate for analysis purposes. For example, EBP data which
contains daily bill-pay transactions for all customers can produce an aggregation
of monthly transactions for each customer per month. These can include count
of transactions, total dollar amount of transactions, average dollar amount of
transactions. If individual transactions had flag values associated with them, then an
aggregate count of flag value occurrences might make sense.

While modeling customer attrition, one of the first steps is to look at periods of
inactivity to determine the appropriate definition of attrition. This is sometimes
referred to as activity analysis. The example analysis below can be extended to
determine that 3 or more consecutive months of inactivity can be considered as
attrition, and customers with more than 25 transactions per month can be classified
as small businesses.

Figure 5: Activity analysis to determine attrition definition

14000 Cummulative % 120%

12000 100%

10000
80%
Frequency

8000
60%
6000
40%
4000

2000 20%

0 0%
0

10

12

14

16

18

20

22

More

Number of inactive months

Figure 6: Activity analysis to identify small business customers

14000 Cummulative % 120%

12000 100%
Count of Customers

10000
80%
8000
60%
6000
40%
4000

2000 20%

0 0%
10

30

50

70

90

110

130

150

170

190

210

230

250

Max # of transactions in a month

Source: Capgemini

6
 the way we see it

4 Cohort and
Trend Analysis

Once a prediction segment has been defined (e.g. attriter or high transactor), the
next step is to look at groups of customers that belong to that segment. In the case
of an attrition model, we can identify customers who attrited in each month and
bucket them into a cohort. For example, JAN09 cohort would be customers whose
three consecutive months of inactivity started in January 2009. This approach leads
to a cohort for nearly every month of data in consideration.

It is possible that each cohort is different – i.e. customers who attrited in one month
exhibit different behavior than customers who attrited in another month. Unless
there are seasonal effects, it is usually unlikely that cohorts are significantly different
from each other. To confirm this, one can compare some attributes of attriters and
non-attriters from different cohorts.

In the example below, average monthly transaction counts of attriters and non-
attriters are plotted for 12 months prior to month of attrition for the cohort. The
four months chosen are Jul 2008, Jan 2009, Jul 09 and Sep 2009.

Figure 7: Cohort analysis to compare behavior across cohorts

8 8
JUL 08 JAN 09
7 7
Count of transactions

Count of transactions

6 6

5 5

4 4

3 3

2 2

ATT_FLAG 1 ATT_FLAG 1
1 1
200801 200801 200802 200802 200803 200803 200804 200804 200805 200805 200806 200807 200807 200808 200808 200809 200809 200810 200810 200811 200811 200812

8 8
JUL 09 SEP 09
7 7
Count of transactions

Count of transactions

6 6

5 5

4 4

3 3

2 2

ATT_FLAG 1 ATT_FLAG 1
1 1
200901 200901 200902 200902 200903 200903 200904 200904 200905 200905 200906 200903 200903 200904 200904 200905 200905 200906 200906 200907 200907 200908

Source: Capgemini

The plots indicate that there is no significant difference between cohorts – whether
it is across years or across months. In each case, there is a difference in level of
activity between attriters and non-attriters. Also, attriters tend to show declining
activity in months close to attrition. These patterns are consistent across all cohorts.

Predictive Modeling Using Transactional Data 7

 These observations allow one to combine all cohorts into one single large segment
of attriters. While combining cohorts, care has to be taken so that monthly activities
are tagged correctly with respect to the month of attrition. If a customer attrited in
Jul 2009, his activity in Jun 2009 will be tagged T-1 and activity in May 2009 will
be tagged T-2. Similarly, for someone who attrited in Jan 2009, activity in Dec 2008
will be tagged T-1 and activity in Nov 2008 will be tagged T-2. Once these tags are
in place, all activity in T-1, T-2 and so on can be aggregated across cohorts.

For example, in the first diagram below, JAN09 cohort had 98 attriters, FEB09
cohort had 105 attriters and so on. Each cohort has 12 months of history that is
considered for analysis. When aggregated, the cohorts stack up as shown in the
bottom diagram.

Figure 8: Aggregating across cohorts

2008 2009

JAN FEB MAR APR MAY JUN JUL AUG SEP OCT NOV DEC JAN FEB MAR APR MAY JUN JUL AUG SEP OCT NOV

98

105

94

97

93

121

117

103

107

T-12 T-11 T-10 T-9 T-8 T-7 T-6 T-5 T-4 T-3 T-2 T-1 T(ATT)

JAN

FEB

MAR

APR
2009

MAY

JUN

JUL

AUG

SEP

8
 the way we see it

5 Model Variable
Definition

Once cohorts are analyzed and combined (if appropriate), the next important step is
to define the set of variables that will be used for modeling.

One obvious set of variables are those associated with the customer and not with
the transactions. These are demographic type of information like Gender, Age,
Location, Marital Status etc. They fluctuate very little over time (except age, which
steadily increases) and are sometimes referred to as stock variables.

While dealing with transactional data, it is useful to look at trends to identify

patterns of customer behavior across time, as shown in the cohort analysis section.
Such attributes are often referred to as time-varying attributes or flow variables.
Since flow variables exhibit high volatility, they are typically aggregated rather than
used as is in models.

Linear trends in flow variables can be captured using two types of variables – one to
capture the level of activity (sometimes referred to as intercept) and one to capture
the trend itself (sometimes referred to as slope). Below is a summary of the types of
variable and the analysis performed on each one.

Figure 9: Types of variables and analysis

Used for
Variable Type Description Type of Analysis Example
Modeling
Stock Variable Static value for Distribution Age YES
customer during
the analysis period

Flow Variable Value changes Time Series Monthly NO

month-to-month (Trend) Transaction
Count

Aggregated Capture intercept Distribution Average YES

Flow Variable and slope of flow Monthly
variable trend Transaction
■ Average Count
■ Average (M-M
Difference)
■ Average of
Directional Flag
■ M12 – M1
■ M9 – M1

Predictive Modeling Using Transactional Data 9

 6 Model Selection
Once model variables are defined, various kinds of models can be created. The most
common ones for predicting customer behavior are logistic regression and decision
trees. These models can be easily created using a tool like SAS or SPSS.

Logistic regression model is created to predict the probability of occurrence of an

event (like attrition) by fitting data into a logistic curve. It uses many predictor
variables that may be numerical or categorical. In case of customer attrition,
variables may reflect the amount of fees paid by the customer in recent months or
change in status of the customer.

Decision trees use tree-like graph or model of decisions to determine the conditional
probability of an outcome (like attrition). It also uses numerical and categorical
variables similar to logistic regression.

Since there are many possible predictive models to choose from, it is useful to have
metrics to compare models and select the best one. Some commonly used metrics
are Receiver Operating Characteristics (ROC) curve, Cumulative Gains Chart and
Lift Chart. All of these provide metrics by trading off desirable outcomes (i.e. correct
predictions) against undesirable outcomes (false positives or false negatives). These
metrics are obtained by running the model on the training data set (used to create
the model) or on an out-of-sample validation set.

ROC Curve plots True Positives along the y-axis and False Positives along the
x-axis. Visually, the higher the curve is above the 45 degree line, and the closer it is
to the top left corner, the better the model.

Figure 10: Sample ROC Curves

Accuracy Profile Plot for Customer Attrition Models

1

0.9

0.8
Cumulative % correct precition

0.7

0.6

0.5

0.4

0.3

0.2

0.1

0
0 100 200 300 400 500 600 700 800 900 1000
Customers ordered by attrition loading
Source: Capgemini

Cumulative Gains Chart and Lift Charts are more commonly used by marketing
departments as they allow for direct visual comparison and interpretation of results
with respect to marketing campaigns.

10
 the way we see it

Cumulative Gains Chart shows the cumulative percentage of target segment

captured (on y-axis) by increasing the number of campaign audience (on x-axis).
This curve typically shows that beyond a certain percentage, additional expansion of
marketing produces little additional benefit in terms of target capture.

Figure 11: Sample Cumulative Gains Chart

1.0

0.8

Cumulative
0.6

0.4

0.2

0.0
0% 20% 40% 60% 80% 100%
Cumulative Baseline
Source: Capgemini

Lift chart directly shows the gain of using the model versus no-model approach.
For example, in the figure above the model works 10 times better when a small
percentage of audience is selected. The effectiveness decreases as the audience widens.

Figure 12: Sample Lift Chart

12

10

8
Lift

0
0% 20% 40% 60% 80% 100%
Lift Lift base
Source: Capgemini

7 Conclusion
Predictive modeling offers the potential for firms to be proactive rather than
reactive. Predictive modeling using transactional data poses particular challenges
which need to be carefully addressed to create useful models. Some of the key
issues identified in this paper are data quality, cohort and trend analysis, model
variable definition and model selection.

Predictive Modeling Using Transactional Data 11

 www.capgemini.com/financialservices

About Capgemini and the

Collaborative Business Experience

Capgemini, one of focused methods and tools. Capgemini

the world’s foremost
providers of Consulting, Technology
and Outsourcing services, has a unique
way of working with its clients, called
the Collaborative Business Experience.
utilizes a global delivery model called
Rightshore® which aims to offer the
right resources in the right location at
competitive cost, helping businesses
thrive through the power of collaboration.

Backed by over three decades of industry
and service experience, the Collaborative
Business Experience™ is designed to
help our clients achieve better, faster,
more sustainable results through seamless
access to our network of world-leading
technology partners and collaboration-
Capgemini reported 2009 global revenues
of EUR 8.4 billion and employs over
90,000 people worldwide.
More information about our services,
offices and research is available at
www.capgemini.com.

About Capgemini and the

Collaborative Business Experience
Copyright © 2010 Capgemini. All rights reserved.
Capgemini, one of focused methods and tools. Capgemini
the world’s foremost utilizes a global delivery model called
providers of Consulting, Technology Rightshore® which aims to offer the
and Outsourcing services, has a unique right resources in the right location at
way of working with its clients, called competitive cost, helping businesses
the Collaborative Business Experience. thrive through the power of collaboration.

Backed by over three decades of industry
and service experience, the Collaborative
Business Experience™ is designed to
help our clients achieve better, faster,
more sustainable results through seamless
access to our network of world-leading
technology partners and collaboration-
Capgemini reported 2009 global revenues
of EUR 8.4 billion and employs over
90,000 people worldwide.
More information about our services,
offices and research is available at
www.capgemini.com.

Copyright © 2010 Capgemini. All rights reserved.