By John J. Coffey, C.P.A. and Gene Palm, www.profitres.com, originally published September 2002, ABA Bank Marketing

§         The purpose of modeling

§         Types of modeling

§         Uses of modeling

There are three things that do not naturally mix well:

1.       Oil and water

2.       Gasoline fumes and an open flame

3.       Bank marketing and statistical modeling!

However, when something causes these things to mix, powerful forces are unleashed.  When detergent is added to water, it can wash stubborn oil stains out of your clothing.  When gasoline fumes are contained in an engine, an open flame from a spark plug can power your car.  When user-friendly modeling software gets into the hands of bank marketers, it can enable them to become enormously effective in their marketing efforts!

The purpose of modeling

This month we are pleased to feature the work of Richard Li, Ph.D., President, Demographic Research Company, an expert in the development of advanced modeling software for bank marketing applications.

According to Richard Li, “The main purpose of modeling data is to optimize the targeting and selection of prospects and customers for direct marketing applications.”  In other words, modeling can enable you to cherry pick customers from your MCIF for direct mail or telemarketing promotions.  Targeting only the best candidates will improve response rates, increase average revenue per piece mailed or per call made and improve the profitability of marketing programs.

In its simplest form, modeling is a lot like profiling.  Profiling is nothing more than examining a group of customers with certain characteristics (e.g., product use, demographics, etc.), constructing a profile from this data, and finding other customers who look like the profile.

For instance, you may have a highly profitable group of customers that have high balances in their money market account and also have a mortgage with you.  Profiling these customers could involve researching their money market balance, age, income and home value.  Then, you would construct a profile of desired characteristics and look for other customers who had similar characteristics, like high balances in their money market accounts who also had the same age, income and home values as the profile, but did not have a mortgage with your bank.

Modeling extends the results of this simple process by taking it one further step.  While modeling reveals the identifying characteristics in a profile, it also predicts and computes a numerical score that reflects the degree of resemblance between the profile and a targeted customer.  By scoring the customers, modeling will enable you to differentiate your customers by the scores and partition them into segments of unequal quality and unequal performance potential.

Types of modeling

There are many types of modeling but we’ll tackle three of the most common modeling methods on the market today – Regression Analysis, Neural Network Analysis and CHAID analysis.

Regression Analysis, such as linear regression, has served as the workhorse of the modeling industry.  For predicting a dependent variable such as propensity to have a mortgage, regression analysis fits a linear equation to the data, which is defined by a combination of independent variables (e.g., money market balance, age, income and home value).  Fitting the equation to the data is optimized when the difference between the actual data (e.g., having a mortgage) and the predicted values is minimized.  A regression equation predicts how the dependent variable will change when you change the values of one or more of the independent variables.

Neural Network Analysis, such as a back propagation neural net, is an associative learning algorithm.  It simulates how our brains learn and establish relationships with data from a learning sample.  On a record-by-record basis, it examines and determines the relationships between a set of output and a set of input variables in the learning set.  Neural network computation begins by assigning initial estimates of the parameter values linking all input variables to output variables.  It then successively adjusts the initial values with each new record examined.  Adjustments are made to minimize the observed difference between the outputs predicted by the neural net’s assigned parameter estimates and the actual data, in the learning set.   Adjustments terminate and the optimum neural network solution is obtained when the observed difference between the actual and the predicted outputs is minimized.  The optimum neural network solution and parameters are found when the difference between the actual and the predicted amount is less than a preset error tolerance level.  Neural network analysis is highly suited to capturing non-linear relationships in data.

CHAID Analysis, which stands for “Chi-Squared Automated Interaction Detection” constructs tree-like hierarchical relationships among variables.  A CHAID tree typically displays many extended branches.  Each branch represents a particular combination of variables in a nested structure.  For example, at the top of the hierarchy may be a specific balance amount for a customer’s money market account.  Nested within this branch may be several sub-branches representing various age groupings.  Each of these age groups can be further sub-divided into several additional income sub-branches and each income sub-branch could be broken into several other home value sub-branches.  Tracing down the successive nodes of each branch, a CHAID model will reveal the combinations of variables (e.g., age, income and home value) that drive different levels of propensity to have a mortgage.  And among those, it will reveal the combination of characteristics that drive the most likely segment of customers.

Uses of modeling

Modeling can be used for just about anything – it can predict potential profit and identify high credit risk customers or prospects.  It can be used to aid customer retention by identifying indicators or triggers of a customer’s likelihood to close an account.  And, it can segment and target customers most likely to purchase the next product based on the nature of a prior purchase.

In short, statistical modeling can unleash the powerful forces of database marketing for impressive results!

John J. Coffey, C.P.A. and Gene Palm are the principals of Profit Resources, a consulting company that specializes in MCIF technologies.  © Profit Resources, Inc. 2006

Richard P. Li, Ph.D. is President of Demographic Research Company, a list segmentation and predictive modeling software firm in the direct marketing industry.  www.drcmodel.com