Explainable Machine Learning for Loan-Delinquency Prediction in U.S. Community Banks: A Quantitative Study Using Gradient Boosting and SHAP

Abstract

Community banks hold a large share of U.S. small-business and local consumer credit, yet their loan portfolios are small, their delinquency events rare, and their supervisory obligations to explain credit decisions exacting. This study asks whether gradient-boosted decision trees can improve 12-month delinquency prediction over a transparent logistic regression in such a setting, and whether Shapley additive explanations (SHAP) can make the resulting model auditable under model-risk and fair-lending expectations. Because loan-level data from community banks are not publicly available, the study constructs a fully documented synthetic panel of 24 banks, about 39,000 loans and 218,573 quarterly loan snapshots (2018Q1–2023Q2) whose generating equations encode non-linear score effects, affordability thresholds, loan-type-specific collateral effects, a county-unemployment shock and deposit-relationship signals, with a 12-month 90-days-past-due label. Logistic regression, random forest, XGBoost, LightGBM and a monotone-constrained LightGBM were tuned on a validation period and evaluated out of time after a four-quarter label gap. The monotone LightGBM achieved the best discrimination (AUC 0.910, 95% CI [0.903, 0.918]; average precision 0.321) against 0.899 and 0.274 for logistic regression; the AUC difference was significant by DeLong's test (z = 6.0) and across six rolling-origin folds (corrected resampled t = 3.29, p = .022), and the constraints cost nothing relative to the unconstrained model. Platt scaling reduced the final model's expected calibration error from 0.033 to 0.010. SHAP attributions ranked current credit score, time since last delinquency, origination score, payment-to-income ratio and loan age highest, were stable across folds (mean Spearman ρ = 0.88) and agreed moderately with logistic coefficients (ρ = 0.71). Deposit-relationship signals added incremental value beyond bureau and payment data. Segment audits showed higher flag and false-positive rates for younger, lower-income and thin-file borrowers, and quarterly monitoring located a discrimination dip after the 2020 shock. A cost analysis indicated that model-ranked outreach to the top decile would avert more loss than a rule-based watch list. Results describe the synthetic portfolio only; the paper's contribution is a replicable, governance-aligned protocol for community banks.