Regulatory Arbitrage or Random Errors? Implications of Race Prediction Algorithms in Fair Lending Analysis

Proxies for race are commonly used in settings where race cannot be observed directly. In the context of small business lending, we examine the standard race prediction algorithm (BISG), which regulators use to assess compliance with fair lending laws. The algorithm relies on an individual’s name and geographical location. If these features are correlated with socioeconomic characteristics, BISG errors could bias fair lending assessments and incentivize lenders to manipulate who they serve, specifically to lend to non-Black borrowers who are falsely predicted to be Black by BISG. We explore these issues using two datasets: proprietary loan application data from an online small business loan marketplace and loan data from the Paycheck Protection Program. We develop a measure of perceived race using images, which we show better correlates with self-identified race than BISG. BISG poorly predicts whether an individual is Black, generating more false classifications than correct ones, and these errors are systematically related to measures of socioeconomic advantage. For example, BISG has especially high false positive rates when classifying Black applicants in areas with high racial animus, where fair lending evaluation may be most critical. In a horse race, image-based race predicts loan approval, while BISG-based race does not, showing that BISG fails to capture important characteristics linked to race that are observable to lenders. There is large variation across lenders in the rate at which they lend to individuals who BISG erroneously assigns to the wrong racial group, leading them to appear more or less compliant with fair lending rules than they would using image-based race. Overall, our study documents the systematic biases in race proxies that rely on name and geography and highlights their implications for racial disparities in lending