Deep Heckman for Loan Evaluation

Despite strong predictive performance, deep learning models for loan classification can suffer from systematic bias when trained on non-representative samples that exclude rejected applicants, as repayment outcomes are only observed for funded applications. We extend the classical Heckman correction method to deep learning architectures by jointly modeling loan selection and repayment outcomes along with their correlation to address unobserved confounding factors. Using a peer-to-peer lending dataset, we demonstrate that our Heckman-corrected deep learning model significantly outperforms benchmark methods, improving both prediction accuracy and generalizability across the complete population of loan applications.