Use of validation on historical data to evaluate fraud models

January 13, 2010 by Chris Ryan

Conducting a validation on historical data is a good way to evaluate fraud models; however, fraud best practices dictate that a proper validation uses properly defined fraud tags.

Before you can determine if a fraud model or fraud analytics tool would have helped minimize fraud losses, you need to know what you are looking for in this category.  Many organizations have difficulty differentiating credit losses from fraud losses.  Usually, fraud losses end up lumped-in with credit losses. When this happens, the analysis either has too few “known frauds” to create a business case for change, or the analysis includes a large target population of credit losses that result in poor results.

By planning carefully, you can avoid this pitfall and ensure that your validation gives you the best chance to improve your business and minimize fraud losses.

As a fraud best practice for validations, consider using a target population that errs on the side of including credit losses; however, be sure to include additional variables in your sample that will allow you and your fraud analytics provider to apply various segmentations to the results.  Suggested elements to include in your sample are; delinquency status, first delinquency date, date of last valid payment, date of last bad  payment and indicator of whether the account was reviewed for fraud prior to booking.

Starting with a larger population, and giving yourself the flexibility to narrow the target later will help you see the full value of the solutions you evaluate and reduce the likelihood of having to do an analysis over again.