Customer:

Internet fraud is ubiquitous, forcing today’s successful companies to go to great lengths to protect themselves from illegal behavior.  But what do you do when your business is entirely internet-based and your industry is a global fraud target?

Such was the case – and the challenge – for leading sub-prime, internet-based lender GECC Finance.  GECC is one of the top lenders of its kind, reviewing up to 80,000 loan applications per day, all of which are received, processed, and approved (or declined) via the internet.

Problem:

Lenders of any type, internet-based or otherwise, make money by loaning out their assets.  If these assets are underutilized, meaning that the lender is not making as many loans as they could, then there are lost revenue opportunities.  However, loaning to the wrong people, namely those that ultimately default, has the obvious effect on revenue.  Not only is the interest revenue lost, but the principle is lost as well, impacting the lender’s ability to make future loans.

Add to this conundrum to the fact that this particular industry’s standard default rate is a staggering 20%.  Even the industry’s best scoring model, at a price of $4M, carries an expected default rate of 13%.

GECC’s goal was to beat the industry average default rate, yet increase their asset utilization (i.e., make more loans).

Methodology:

For this analysis, GECC extracted historical records of one hundred thousand transactions, both good and bad, and augmented this data with publicly available demographic data.  With minimal preprocessing or data manipulation, GECC’s analyst built 100 random trees on a training data set of 20,000 records.

Interestingly, the customer was a new Optimus RP user, and this entire analysis was performed as part of a two-hour web-based training exercise.

Results:

The resultant model, built in less than one day, created business rules with a predicted default rate of only 10%, yet also allowed GECC to loosen certain credit restrictions, enabling them to ultimately approve a greater number of loans.  The resultant model has performed as designed in actual use, resulting in increased revenues for GECC, coupled with a lower rate of default.