Explainable Machine Learning for Corporate Bankruptcy Prediction: Evidence from Accounting Information

Bankruptcy prediction sits at the core of credit risk management, yet the ML models now used for it are seldom examined for interpretability. This paper benchmarks five classifiers — Logistic Regression, XGBoost, Support Vector Machine, K-Nearest Neighbors, and a Neural Network — on 78,682 firm-year records of U.S. public companies from 1999 to 2018. Boruta selection screens 36 raw items and accounting ratios down to the informative ones, and class imbalance is tackled with cost-sensitive learning plus a Youden's J threshold. XGBoost attains the highest test AUC (0.8343), trailed by ANN (0.8255) and KNN (0.8149). Retained earnings ratio, market-to-assets, and EBIT emerge from SHAP as the strongest predictors. Across pre-crisis (1999–2006), crisis (2007–2009), and post-crisis (2010–2011) windows, the relative importance of these indicators stays largely unchanged, with retained earnings leading in every period. By pairing strong predictive accuracy with explainable evidence, the study narrows the gap between black-box performance and the transparency expected by lenders and regulators. The resulting framework acts as an auditable early-warning tool that practitioners can both trust and interrogate.