Modeling corporate financial distress using explainable machine learning: evidence from macro-financial panel data

Purpose This study aims to examine whether firm-level market indicators and lagged macroeconomic information can be integrated within an explainable predictive framework for corporate financial distress, with particular attention to temporal dynamics, model performance and interpretability. Design/methodology/approach This study uses an unbalanced panel of 15,664 firm-month observations from 99 large-cap US listed equity series covering July 2012–September 2025. Regularized logistic regression, Random Forest, HistGradientBoosting and XGBoost are evaluated using chronological training, validation and held-out testing. SHapley Additive exPlanations (SHAP) provide model interpretability, while alternative macroeconomic lag structures, calibration diagnostics and robustness checks assess the stability of the predictive evidence. Findings Regularized logistic regression records the strongest full-test point estimates (AUC = 0.7886; average precision = 0.3085), while XGBoost is the strongest nonlinear benchmark (AUC = 0.7314; average precision = 0.2049). However, model rankings vary across evaluation periods. The combined one-, three- and six-month lag specification performs best, indicating that predictive macroeconomic information is distributed across multiple horizons. SHAP identifies 63-day firm-level volatility as the leading XGBoost feature and reveals conditional relationships between macroeconomic conditions and firm-specific vulnerability. Research limitations/implications The large-cap US sample, market-based distress definition and mixed-frequency design limit broader generalization. Practical implications The framework can support risk ranking, early-warning screening and interpretable financial-risk monitoring for investors, financial institutions and regulators. Originality/value This study integrates temporal macroeconomic information, comparative model evaluation and explainable AI, demonstrating that greater model complexity does not necessarily improve out-of-sample distress prediction.