Predicting Financial Distress: A Comparative Analysis of Explainable Machine Learning Models for Cross‐Market Bankruptcy Prediction

Abstract

Despite substantial advances in machine‐learning approaches to bankruptcy prediction, less is known about whether explainable models remain reliable across markets and during structural breaks. This study examines that issue by comparing two explainable architectures, TabNet–SHAP and NGBoost–SHAP, for 1‐year‐ahead bankruptcy prediction in the United States and China during 2015–2022. Using a balanced panel of listed firms from both markets, we evaluate out‐of‐sample performance before and after COVID‐19 and benchmark both models against conventional linear and ensemble classifiers. Model uncertainty is assessed through cross‐validation and bootstrap confidence intervals. The results show that NGBoost–SHAP provides more stable and better‐calibrated probabilistic predictions across regimes, whereas TabNet–SHAP produces more variable results and is more sensitive to regime shifts and preprocessing choices. We interpret this pattern as an architecture‐specific trade‐off between stability and adaptability rather than as an unconditional ranking of the two models. SHAP‐based interpretation indicates that profitability and capital‐structure variables account for most of the explanatory signal within each market and model. The post‐COVID Chinese subsample shows the most concentrated feature‐attribution structure, although SHAP magnitudes are interpreted only as within‐model rankings rather than as directly comparable cross‐market quantities. The study contributes by proposing an adaptability–stability–interpretability framework for selecting bankruptcy prediction models under different institutional and market conditions. It also provides evidence that explainable ensemble and deep tabular models can recover economically meaningful distress signals across markets while differing in their calibration, stability, and sensitivity to economic regime change.