Testing Credit Scoring Models for Statistical Bias: Ushering in a New Era of Transparency

August 13, 2018

In the consumer credit market, one of the most important challenges is to ensure credit scoring models are free from what is known as “statistical bias.” Statistical bias concerns about a particular credit score model arise if the probability of default (PD) varies between various groups of consumers even though members of the respective groups receive the same score with the same model.

If there are significant differences in the probability of default between any population segments, the credit score model then has a statistical bias (either positive or negative) towards one or more such consumer groups. This, in turn, suggests that there is preferential treatment or mistreatment of those particular population segments.

This paper examines if the latest VantageScore credit scoring model VantageScore 4.0 (like all other VantageScore models) evenly distributes PD across population segments and provides a methodology for best practices for lenders to test their own portfolios. The examination includes testing ethnic populations as well as new scoring (universe expansion) vs. mainstream scoring consumers.

Popular Articles

Consumer FAQ: Benefits of Adding Rent and Utility Data to a Credit File

Advantage of Adding Rent and Utility Data whitepaper

Credit with a Conscience fact sheet

Driving Financial Inclusion with Data and Analytics fact sheet

Credit Invisible No Longer: Examining the relationship between socioeconomic disparities and scoreability