A subsystem-aware explainable AI delivers high-performing, auditable bankruptcy predictions on a standard Taiwanese benchmark, surfacing leverage, profitability and liquidity as key drivers; but findings rest on a single, small dataset and stop short of claiming universal superiority over traditional tabular learners.

Main Finding

A systems-oriented explainable AI framework (Hierarchical LFRM-MACI) that refines subsystem-specific features and models cross-subsystem interactions can produce high-performing and auditable corporate bankruptcy early-warning predictions on the UCI Taiwanese Bankruptcy dataset. Combined predictive benchmarking and transparent attribution (PFI, SHAP) reveal economically meaningful drivers (leverage, profitability, liquidity, efficiency, growth/shareholder performance), suggesting XAI is practical for risk-management contexts even when not claiming blanket superiority over traditional tabular learners.

Key Points

• Proposed architecture: Hierarchical LFRM-MACI
  • LFRM (Local Feature Refinement Module) learns refined representations within financial subsystems: profitability, solvency, liquidity, efficiency, and growth/shareholder performance.
  • Cross-subsystem attention (MACI) models interactions between these subsystem representations.
• Evaluation strategy:
  • Experiments run under two class-ratio settings: balanced (1:1.0) and moderately imbalanced (1:2.5).
  • Reporting distinguishes traditional ML vs deep learning benchmarks; traditional tabular learners included in a formal single-split comparison, while the main contribution is structured representation + attribution workflow rather than raw algorithmic dominance claims.
• Predictive performance (ROC-AUC):
  • Single 70%/30% split: 0.9286 (1:1.0) and 0.9269 (1:2.5).
  • Repeated 5-fold CV (5 repetitions): 0.8940 [0.8777, 0.9103] (1:1.0) and 0.9135 [0.9001, 0.9269] (1:2.5).
• Interpretability:
  • Permutation Feature Importance (PFI) and SHAP used to produce subsystem-level attributions.
  • Explanation patterns emphasize predictors tied to leverage pressure, profitability and asset structure, liquidity strains, operating efficiency, and growth/shareholder metrics.
• Positioning: The study emphasizes structured deep representations plus transparent, auditable attribution as the practical contribution for financial decision-making rather than claiming outright superiority over all traditional tabular models.

Data & Methods

• Data:
  • Public UCI Taiwanese Bankruptcy Prediction dataset (standard benchmark in corporate bankruptcy literature).
  • Experiments on both balanced and moderately imbalanced class settings to reflect realistic skew.
• Model architecture:
  • Hierarchical design that first refines features within domain-informed subsystems (LFRM) and then applies cross-subsystem attention (MACI) to capture interactions.
  • Framing is systems-oriented: financial conditions are modeled as interacting subsystems rather than flat feature vectors.
• Baselines and benchmarking:
  • Benchmarks separated into traditional machine learning (tabular learners) and deep learning methods.
  • Traditional tabular learners included in single-split empirical comparison; structured deep approach is presented as complementary (structured representation + attribution).
• Evaluation protocol:
  • Single 70/30 train/validation split for formal comparison.
  • Repeated 5-fold cross-validation (5 repetitions) for more robust performance estimates; confidence intervals reported for ROC-AUC.
• Interpretability methods:
  • Permutation Feature Importance (PFI) for subsystem-level importance.
  • SHAP values to attribute contributions of features/subsystems to individual and aggregate predictions.
• Reporting:
  • Emphasis on transparent and auditable attribution workflows as part of decision support.

Implications for AI Economics

• Modeling insight:
  • Treating financial distress as a system of interacting subsystems (profitability, solvency, liquidity, efficiency, growth/shareholder returns) aligns model structure with economic theory and can improve interpretability and trust in predictions.
• Policy and governance:
  • Subsystem-level attributions (PFI, SHAP) support regulatory and audit requirements by making drivers of risk explicit—important for model governance, stress testing, and compliance.
• Risk management applications:
  • The combined predictive + attribution workflow enables actionable risk signals (e.g., identification of leverage or liquidity pressure) that can be integrated into monitoring, lending decisions, and early-intervention policies.
• Methodological trade-offs:
  • Structured deep representations provide richer interaction modeling than flat tabular models, but the paper positions the contribution as adding interpretability and structured representation rather than outperforming all traditional learners in every setting. Small benchmark datasets and evaluation choices (single split vs CV) affect claims of general superiority.
• Research and deployment considerations:
  • Need for external validation on larger and more diverse corpora before deployment; sensitivity to class imbalance and dataset size must be managed (cost-sensitive training, calibration).
  • Attribution methods (PFI, SHAP) are useful but should be combined with robustness checks (e.g., counterfactuals, stability under dataset shifts) for high-stakes financial decisions.
• Economic research opportunities:
  • Applying systems-oriented XAI to other corporate outcomes (default timing, recovery rates), exploring temporal extensions (time-series subsystem dynamics), and integrating macroeconomic shock scenarios for stress testing and policy analysis.