Bankruptcy Prediction via Hybrid Resampling and Stacking Ensemble Techniques with Explainable Artificial Intelligence (XAI)-Driven Analysis

Technical Summary: Bankruptcy Prediction via Hybrid Resampling and Stacking Ensemble Techniques with XAI-Driven Analysis

Problem Statement
The study addresses the critical challenge of predicting corporate bankruptcy within severely imbalanced financial datasets, where the minority class (bankrupt firms) constitutes a small fraction of the total population (approximately 3.23% in the target dataset). Traditional evaluation metrics like accuracy are misleading in such contexts, often favoring the majority class and failing to detect financially distressed firms. The paper argues that bankruptcy prediction is a cost-sensitive decision problem where missing a distressed firm (low recall) carries higher consequences than false alarms. Furthermore, the field requires models that are not only predictive but also interpretable to support auditability, governance, and regulatory compliance.

Methodology
The research employs a comprehensive, end-to-end framework utilizing the Taiwanese Bankruptcy Prediction dataset from the UCI Machine Learning Repository (6,819 firm observations, 95 predictors). The methodology proceeds through four distinct stages:

1. Consensus-Based Feature Selection: To reduce the high-dimensional input space (95 features) and enhance robustness, five feature selection algorithms were applied: Whale Optimization (WOA), Particle Swarm Optimization (PSO), Salp Swarm Optimization (SSO), Bat Algorithm (BA), and Mutual Information (MI). A consensus retention rule was applied, retaining only predictors identified by at least two of the five algorithms. This reduced the feature set to 23 robust variables, including metrics related to profitability, leverage, solvency, and operational efficiency.
2. Hybrid Resampling: To address class imbalance, three hybrid resampling techniques were implemented on the training data: SVM-SMOTE, SMOTE-Tomek, and SMOTE-Edited Nearest Neighbors (SMOTE-ENN). These methods were used to generate balanced training distributions while attempting to preserve the integrity of the minority class.
3. Model Architecture and Stacking:
   • Base Learners: Five ensemble machine learning classifiers were evaluated: Gradient Boosting (GB), Extreme Gradient Boosting (XGB), Histogram-based Gradient Boosting (HGB), LightGBM (LGBM), and AdaBoost (AB).
   • Meta-Learners: Five deep learning (DL) models were tested: Recurrent Neural Network (RNN), Long Short-Term Memory (LSTM), Gated Recurrent Unit (GRU), Deep Neural Network (DNN), and Multi-Layer Perceptron (MLP).
   • Hybrid Stacking: A stacking ensemble framework was constructed where the five ML classifiers served as base learners, and each of the five DL models served as a meta-learner, creating 25 unique hybrid configurations.
4. Explainable AI (XAI): The Shapley Additive exPlanations (SHAP) framework was applied to the best-performing models to interpret feature contributions at both global and local levels, ensuring the model's decision logic aligns with economic theory.

Key Results
The study evaluated models using accuracy, recall, specificity, G-mean, and ROC-AUC. Key findings include:

• Impact of Resampling: The choice of resampling strategy significantly dictated model behavior. SVM-SMOTE and SMOTE-Tomek favored high accuracy and specificity (correctly identifying non-bankrupt firms), whereas SMOTE-ENN yielded superior minority-class detection (recall) and balanced performance (G-mean).
• Standalone Model Performance: Among standalone models, the GRU classifier combined with SMOTE-ENN achieved the best overall predictive balance, recording a recall of 0.8627, a G-mean of 0.8517, and a ROC-AUC of 0.9431.
• Stacking Ensemble Performance: The hybrid stacking ensemble using SMOTE-ENN with (GB+XGB+HGB+LGBM+AB) as base learners and LSTM as the meta-learner provided the strongest compromise between sensitivity and specificity among the stacked models (Recall: 0.7255, G-mean: 0.8254, ROC-AUC: 0.9270).
• Model Comparison: While tree-based ensembles (specifically LGBM) excelled in accuracy and specificity, deep learning models (MLP, GRU, LSTM, RNN) consistently outperformed machine learning models in recall and G-mean, particularly under the SMOTE-ENN regime.
• Feature Importance: SHAP analysis identified leverage, profitability, solvency, and operational efficiency indicators (e.g., Retained Earnings to Total Assets, ROA, Debt Ratio, Total Asset Turnover) as the most influential predictors of bankruptcy risk.

Significance and Claims
The paper claims that its primary contribution lies in demonstrating that class imbalance is a central determinant of model utility, not merely a preprocessing nuisance. The study asserts that:

1. Deep Learning Viability: Deep learning architectures (specifically GRU and LSTM) offer substantial gains in detecting the minority bankrupt class even in structured tabular settings, outperforming traditional boosting methods in recall.
2. Resampling Strategy Matters: There is no single "best" resampling method; the optimal choice depends on the operational objective. SMOTE-ENN is shown to be particularly effective for early warning systems where minimizing missed bankruptcies is the priority.
3. Interpretability Integration: By integrating SHAP, the study validates that the models capture economically plausible signals (profitability and leverage) rather than statistical artifacts, thereby supporting the deployment of these models in high-stakes financial governance.
4. Benchmarking: The study establishes a robust benchmark using structured financial ratios alone, suggesting that when combined with consensus feature selection, hybrid resampling, and XAI, traditional accounting data remains highly informative for bankruptcy prediction.

The authors conclude that while their framework advances the field, future work should focus on integrating multi-period trajectories and unstructured textual data to further enhance robustness and generalizability across different regulatory and economic environments.