Machine Learning, Alternative Data, and Mortgage Bias

This paper evaluates the impact of machine learning (ML) and alternative data (AD) on ethnic/racial bias in residential mortgage underwriting. Based on the newly enhanced Home Mortgage Disclosure Act data, we simulate data bias and AD bias and construct multiple test scenarios by interacting the two sources of bias. The distributional impact of ML is compared with conventional statistical models across test scenarios. We find that bias in historical data could reinforce itself in future decisions when decisioning algorithms leverage biased AD features to proxy for prohibited ethnicity/race group memberships. Moreover, while AD features linearly correlated with minorities affect both ML and traditional statistical models, ML models can result in larger bias by uncovering sophisticated ethnicity/race proxies invisible to statistical models