Can Artificial Intelligence Improve Credit Decisions? Evidence from Large-Scale Lending Data

Abstract

This study examines the use of machine learning for predicting default risk within the U.S. Small Business Administration (SBA) 7(a) loan program using publicly available loan-level data obtained through the Freedom of Information Act (FOIA). Using 488,772 resolved loans originated over fifteen fiscal years from FY2010 through FY2024, the study compares a traditional logistic regression benchmark with Random Forest, Gradient Boosting, and an ensemble model. Gradient Boosting produces the strongest out-of-time predictive performance, achieving an area under the receiver operating characteristic curve (AUC-ROC) of 0.9178 compared with 0.7641 for the logistic regression benchmark, representing a 20.1% relative improvement. The model also reduces the Brier Score by 44.5%, indicating substantially more accurate probability forecasts. Loan term, initial interest rate, and the SBA guarantee ratio emerge as important predictors of default. Default outcomes also vary considerably across industries, with Transportation (9.84%), Arts and Entertainment (9.70%), and Wholesale Trade (8.52%) exhibiting the highest default rates, while Agriculture (4.09%), Finance and Insurance (4.60%), and Healthcare (4.81%) exhibit the lowest. Across the study period, among the resolved-loan sample has an overall default rate of 7.47%, with annual rates ranging from 5.48% in FY2013 to 16.27% in FY2023. The findings demonstrate the potential value of machine learning for credit risk prediction in government-guaranteed small business lending while highlighting the importance of out-of-time model validation. More broadly, the study provides evidence relevant to SBA policymakers, participating lenders, and the responsible development of AI-assisted credit risk assessment tools for small business lending.