One threshold doesn't fit all: Tailoring machine learning predictions of consumer default for lower‐income areas

Abstract

Improving fairness across policy domains often comes at a cost. However, as machine learning (ML) advances lead to more accurate predictive models in fields like lending, education, healthcare, and criminal justice, policymakers may find themselves better positioned to implement effective fairness measures. Using credit bureau data and ML, we show that setting different lending thresholds for low‐ and moderate‐income (LMI) neighborhoods relative to non‐LMI neighborhoods can equalize the rate at which equally creditworthy borrowers receive credit. ML models alone better identify creditworthy individuals in all groups but remain more accurate for the majority group. A policy that equalizes access via separate thresholds imposes a cost on lenders, but this cost is outweighed by the substantial gains from ML. This approach aligns with the motivation behind existing laws such as the Community Reinvestment Act, which encourages lenders to meet the credit needs of underserved communities. Targeted Special Purpose Credit Programs could provide the opportunity to prototype and test these ideas in the field.