Predictably Unequal? The Effects of Machine Learning on Credit Markets

Innovations in statistical technology have sparked concerns about distributional impacts across categories such as race and gender. Theoretically, as statistical technology improves, distributional consequences depend on how changes in functional forms interact with cross-category distributions of observable characteristics. Using detailed administrative data on US mortgages, we embed the predictions of traditional logit and more sophisticated machine-learning default prediction models into a simple equilibrium credit model. Machine learning models slightly increase credit provision overall, but increase rate disparity between and within groups; effects mainly arise from flexibility to uncover structural relationships between default and observables, rather than from triangulation of excluded characteristics. We predict that Black and Hispanic borrowers are disproportionately less likely to gain from new technology