Predictably Unequal? The Effects of Machine Learning on Credit Markets

Abstract

Recent innovations in statistical technology, including in evaluating creditworthiness, have sparked concerns about impacts on the fairness of outcomes across categories such as race and gender. We build a simple equilibrium model of credit provision in which to evaluate such impacts. We find that as statistical technology changes, the effects on disparity depend on a combination of the changes in the functional form used to evaluate creditworthiness using underlying borrower characteristics and the cross-category distribution of these characteristics. Employing detailed data on US mortgages and applications, we predict default using a number of popular machine learning techniques, and embed these techniques in our equilibrium model to analyze both extensive margin (exclusion) and intensive margin (rates) impacts on disparity. We propose a basic measure of cross-category disparity, and find that the machine learning models perform worse on this measure than logit models, especially on the intensive margin. We discuss the implications of our findings for mortgage policy.