Problem:

Attracting and acquiring new customers is important and expensive for any business, but keeping the ones you’ve got is arguably more so.  Nowhere is this more true, and perhaps more difficult, than in the world of banking and financial services where customer loyalty is transient, and the incentives to change are many.
This challenge is well known, and most financial services institutions take extraordinary care to prevent defection.  Thus predictive modeling for attrition is a key focus.

For some time, this mid-sized American bank had been working on and refining a model to predict which of their best customers were at risk of leaving, based on account and transaction attributes.  In this particular example, of particular interest were the top few percent of at-risk customers.  This bank is sophisticated in predictive analytics, with various tools being used by a number of highly experienced analysts. 

After working on the project for some time with TreeNet and other tree-based tools, these modelers had been able to achieve a 91.3% level of accuracy in predicting attrition of the top 1% of the customers.

Methodology:

For this analysis, the bank extracted various pieces of information from 108,000 accounts and also on 84,000 customers.  In the analysis process, these two tables were joined within Optimus RP.  There were no other transformations and no additional covariates were created or augmented.  The data was randomly split into training and test sets based on 50% of the real data.  1,000 random trees were automatically created with an adjusted p-value cut-off of 0.01.

Total modeling time was approximately one hour, plus computer run time of 2 hours.

Results:

The bank judged the Optimus RP results, reporting that of the top 1% of customers predicted most likely to defect, 98.1% actually did – a  nearly 7% gain over earlier attempts, and with a fraction of the modeling time.