Conference Agenda

We study fairness-aware credit modeling with interpretable machine learning in U.S. mortgage markets. We introduce an in-processing training objective that augments the classification error with a differentiable equalized-odds penalty, ensuring parity in error rates across protected groups. Our fairness-aware machine learning model provides inherent interpretability, allowing for the decomposition of model outputs into feature contributions. Using HMDA data, we find that the fairness-regularized model substantially narrows TPR/FPR disparities and shifts importance from proxy-like variables toward core underwriting determinants, yielding transparent, regulator-aligned decisions. The resulting models constitute realistic, interpretable less-discriminatory alternatives to standard credit scoring rules, showing that substantial reductions in disparate treatment can be achieved at minimal cost. Quasi-experimental tests around underwriting thresholds and lender-level comparisons confirm that fairness regularization lowers minority shortfalls at the margin and reduces within-lender disparities.