Probabilistic boosting outperforms deep tabular nets for stable credit-risk probabilities: NGBoost–SHAP gives better-calibrated, less regime-sensitive 1‑year bankruptcy forecasts across US and China, whereas TabNet–SHAP adapts faster but is more variable and preprocessing-sensitive.

Main Finding

NGBoost–SHAP yields more stable and better‑calibrated 1‑year bankruptcy probabilities across the U.S. and Chinese markets and across the COVID regime break, whereas TabNet–SHAP is more adaptive but produces more variable, regime‑sensitive results. This pattern reflects an architecture‑specific trade‑off between stability and adaptability rather than an absolute ranking of the two models.

Key Points

• Models compared: TabNet–SHAP (deep tabular architecture with SHAP explanations) and NGBoost–SHAP (probabilistic boosting with SHAP).
• Scope: 1‑year‑ahead bankruptcy prediction for listed firms in the United States and China, 2015–2022, using a balanced panel.
• Evaluation: out‑of‑sample performance before and after the COVID shock, benchmarked against conventional linear and ensemble classifiers.
• Uncertainty quantification: cross‑validation and bootstrap confidence intervals used to assess model uncertainty and stability.
• Main empirical results:
  • NGBoost–SHAP produced more stable predictions across regimes and better probabilistic calibration.
  • TabNet–SHAP showed greater variability, higher sensitivity to regime shifts and preprocessing choices, and less stable calibration.
• Explainability findings (SHAP):
  • Profitability and capital‑structure variables carry most of the explanatory signal in both markets and models.
  • The post‑COVID Chinese subsample exhibited the most concentrated feature‑attribution pattern.
  • SHAP values are treated as within‑model feature rankings and are not directly comparable across models or markets.

Data & Methods

• Data: Balanced panel of publicly listed firms in the U.S. and China covering 2015–2022.
• Prediction target: 1‑year ahead bankruptcy/distress.
• Models evaluated:
  • TabNet with SHAP for local/global feature attributions.
  • NGBoost with SHAP to provide probabilistic forecasts and feature attributions.
  • Conventional linear and ensemble baselines for benchmarking.
• Validation strategy:
  • Out‑of‑sample tests split around the COVID regime break (pre/post).
  • Cross‑validation for model selection and stability checks.
  • Bootstrap procedures to compute confidence intervals for performance metrics and SHAP summaries.
• Interpretability: SHAP used to rank feature importance within each model; magnitudes interpreted cautiously as relative (within‑model) signals.

Implications for AI Economics

• Model selection under regime change:
  • Use NGBoost‑like probabilistic ensembles when stability and well‑calibrated risk probabilities are critical (e.g., regulatory stress tests, credit scoring under varying macro states).
  • Consider TabNet‑like deep tabular models for environments where adaptability to new patterns may pay off, but pair them with rigorous validation and robustness checks.
• Model monitoring and governance:
  • Regular out‑of‑sample testing across economic regimes and bootstrap uncertainty quantification are essential to detect calibration drift and preprocessing sensitivity.
  • SHAP summaries are useful to recover economically meaningful drivers (profitability, leverage) but should not be used to compare importance magnitudes across models or markets without caution.
• Policy and risk management:
  • Explainable ML can recover relevant distress signals across institutional contexts, supporting cross‑market credit surveillance, but differing calibration/stability properties imply different operational uses.
• Research agenda:
  • Further work should formalize the proposed adaptability–stability–interpretability framework, study intervention strategies when models disagree, and explore techniques to reduce preprocessing sensitivity in highly adaptive architectures.