The CFPB highlighted the results of an analysis comparing the uses of traditional and non-traditional sources of information by consumers in the credit process.
In 2017, the CFPB granted no-action relief from certain Regulation B requirements to Upstart Network, Inc. (“Upstart Network”) to use alternative data (such as education and employment history) and machine learning for the purpose of an underwriting and pricing model. The no-action letter was contingent on Upstart Network providing the CFPB with information about compared results between (i) its credit underwriting and pricing model (a tested model) and (ii) a more standard model. Upstart Network was tasked with answering:
whether the Alternative Model’s use of alternative data and machine learning would increase access to credit; and
if the Alternative Model’s underwriting or pricing results create greater disparities than the traditional model (i.e., race, ethnicity, sex, age).
Based on the information gathered by Upstart Network, the CFPB found that:
access-to-credit comparisons showed the Alternative Model approved 27 percent more applicants than the traditional model, in addition to yielding 16 percent lower average annual percentage rates (“APRs”) for approved loans;
the expansion of credit access increased the acceptance rates in the Alternative Model for all tested races, ethnicity and sex segments by 23-29 percent while decreasing the average APRs by 15-17 percent;
“near prime” consumers in the Alternative Model with FICO scores between 620 and 660 were approved nearly twice as frequently;
applicants under 25 years of age in the Alternative Model were 32 percent more likely to be approved; and
consumers in the Alternative Model with incomes under $50,000 were 13 percent more likely to be approved.
Should the regulators be approving credit models based on whether they are happy with the results? What happens if another credit scoring metric produces different or less favored results: does that metric become illegal to use without regard to the process of its production or its accuracy?
Big data raises lot of important social/moral questions; and “disparate impact” is one of the more complex ones. For some background discussion of “big tech,” “big data” and credit scoring, see “Big tech in finance: opportunities and risks,” particularly the discussion of credit provision beginning on page 60, and Senate Banking Committee Considers Testimony on Consumer Data Vendors.