Source-linked AI summary
Bankruptcy Prediction via Hybrid Resampling and Stacking Ensemble Techniques with Explainable Artificial Intelligence (XAI)-Driven Analysis
Obu-Amoah Ampomah, Edmund Fosu Agyemang, Kofi Acheampong, Louis Agyekum, Enock Adu Bonsu, Eric Nyarko
TL;DR
Bankruptcy prediction remains difficult because severely imbalanced financial data can obscure minority-class failures and models must balance sensitivity with specificity. The study combines consensus feature selection, hybrid resampling, heterogeneous machine and deep learning, stacking, and SHAP on Taiwanese firm data. Resampling materially changes model behavior: SMOTE-ENN strengthens minority-class detection, while standalone GRU and stacked LSTM models provide the strongest reported predictive compromises, with interpretable risk signals concentrated in leverage, profitability, solvency, and operating efficiency.
Problem
Severely imbalanced bankruptcy data make minority-class detection and reliable evaluation difficult, while existing work comparatively less often combines complementary learners, imbalance-aware metrics, and actionable explanations end to end.
Method
The framework applies consensus feature selection, three hybrid resampling strategies, heterogeneous ML and DL models with stacking, imbalance-aware evaluation, and SHAP explanations to the Taiwanese bankruptcy dataset.
Results
SMOTE-ENN delivered stronger minority-class detection; GRU with SMOTE-ENN achieved ROC-AUC 0.9431, while SMOTE-ENN with (GB+XGB+HGB+LGBM+AB)+LSTM provided the strongest stacking compromise.
Takeaways & Limitations
Structured financial-ratio data remain informative when feature selection, imbalance correction, ensemble learning, and SHAP interpretation are carefully combined.
Takeaways & Limitations
The analysis uses a single Taiwanese dataset and mainly structured financial ratios, limiting generalizability across countries, industries, regulatory environments, and richer information settings.
Abstract
from arXiv · showhide
This study develops and evaluates a bankruptcy prediction framework that integrates consensus-based feature selection, hybrid resampling, stacking ensembles, and explainable artificial intelligence to improve minority-class detection in severely imbalanced financial data. Using the Taiwanese Bankruptcy Prediction dataset from the UCI Machine Learning Repository, five feature-selection algorithms were first applied, and a consensus retention rule reduced the input space to 23 robust variables. The balanced training data were then generated using SVM-SMOTE, SMOTE-Tomek, and SMOTE-ENN. Five ensemble machine learning classifiers, namely gradient boosting, extreme gradient boosting, histogram-based gradient boosting, LightGBM, and AdaBoost, were compared with five deep learning models, including RNN, LSTM, GRU, DNN, and MLP. In addition, hybrid stacking ensembles combined the five machine learning classifiers as base learners with each deep learning model as a meta-learner. Model performance was assessed using accuracy, recall, specificity, G-mean, and ROC-AUC, while SHAP was used to explain feature contributions. The results show that resampling strategy materially shaped model behavior. SVM-SMOTE and SMOTE-Tomek favored accuracy and specificity, whereas SMOTE-ENN delivered stronger minority-class detection. Among standalone models, the GRU with SMOTE-ENN achieved the best overall predictive balance, with recall of 0.8627, G-mean of 0.8517, and ROC-AUC of 0.9431. Among stacking ensembles, SMOTE-ENN with (GB+XGB+HGB+LGBM+AB)+LSTM provided the strongest compromise between sensitivity and specificity. SHAP analysis identified leverage, profitability, solvency, and operational efficiency indicators as the most influential predictors of bankruptcy risk. These findings support more reliable and interpretable early warning systems for financially distressed firms.
1 Introduction
Bankruptcy prediction must address rare failures, shifting financial conditions, and the practical costs of missed distressed firms. The study motivates hybrid, interpretable modeling that combines complementary learners and imbalance-aware evaluation.
- Bankruptcy is a rare-event problem with shifting predictor distributions and evolving firm conditions that static one-period models may not capture.
- Recall, G-mean, AUC, and class-specific error rates are emphasized because missed distressed firms may be more consequential than false alarms.
- Tree-based ensembles capture nonlinear interactions and mixed variable types, while stacking combines base learners to exploit complementary inductive biases.
- Sequence-aware deep learning can outperform conventional baselines when multiple periods of firm information are available and prediction is explicitly temporal.
- Hybrid ensembles are motivated by the absence of a single model family that consistently dominates across imbalance, drift, and sequential challenges.
- Transparency and interpretability support auditability, governance, and human investigation when model outputs trigger consequential actions.
2 Data and Methods
The study uses the Taiwanese Bankruptcy Prediction dataset and a pipeline combining preprocessing, consensus feature selection, hybrid resampling, heterogeneous learners, stacking, and SHAP explanations. Its design evaluates ten models with imbalance-aware metrics while reducing the predictor space through cross-method agreement.
- 2.1 Data Description: The Taiwanese dataset contains 6,819 firm records and 95 financial-ratio predictors collected from 1999 to 2009, with no missing values reported.
- 2.1 Data Description: Bankruptcy cases comprise approximately 3.23% of the dataset, motivating evaluation with recall, ROC-AUC, and other imbalance-aware measures.
- 2.2 Data Preprocessing and method of analysis: Predictors were normalized to [0,1] using a scaler fitted on training data, then applied to both training and test sets.
- 2.4 Feature Selection: Five feature-selection algorithms were combined through a consensus rule retaining predictors identified by at least two methods.
- 2.2 Data Preprocessing and method of analysis: The study used an 80:20 train-test split, three hybrid resampling methods, five ensemble ML base learners, five DL models, and hybrid stacking combinations.
- 2.3 Model Performance Evaluation Metrics: SHAP was applied to estimate global and local feature contributions and identify indicators that increased or decreased bankruptcy classification likelihood.
- 2.4.1 Consensus Based Feature Retention Criterion: Consensus retention reduced the feature space to 23 predictors for subsequent modeling and substantive interpretation.
3 Results and Findings
Results show that resampling strategy systematically shaped the trade-off between majority-class performance and minority-class detection across standalone and stacking models. SMOTE-ENN improved bankruptcy detection most consistently, while SHAP highlighted a concentrated set of financial risk indicators and cross-study comparisons remained limited by dataset heterogeneity.
- Resampling and experimental setup: 6,599 non-bankruptcy cases (96.77%) versus 220 bankruptcy cases (3.23%) defined the original severely imbalanced dataset.Hybrid oversampling was used to improve minority-class representation while retaining majority-class information.
- Standalone model results: Under SVM-SMOTE, LGBM achieved the highest accuracy (0.9545) and specificity (0.9726), while MLP achieved the highest recall (0.8235) and G-mean (0.8574).HGB recorded the highest ROC-AUC (0.9408) among the reported SVM-SMOTE models.
- Standalone model results: Under SMOTE-Tomek, LGBM led accuracy (0.9494) and specificity (0.9673), while MLP achieved the highest recall (0.8627) and ROC-AUC (0.9327).RNN recorded the highest G-mean (0.8481) under this resampling strategy.
- Standalone model results: Under SMOTE-ENN, GRU recorded the best ROC-AUC (0.9431), while LSTM and MLP achieved the highest recall (0.8824) and RNN the highest G-mean (0.8575).LGBM still produced the highest accuracy (0.9289) and specificity (0.9383) among the reported ML models.
- Resampling and experimental setup: SMOTE-SVM produced the strongest average stacking accuracy (mean = 0.9576) and specificity (mean = 0.9785), whereas SMOTE-ENN achieved the best recall (mean = 0.6784) and G-mean (mean = 0.8006).SMOTE-Tomek followed SMOTE-SVM in average accuracy and specificity, while SMOTE-ENN incurred more false positives among non-bankrupt firms.
- Stacking and explainability: SMOTE-ENN+(GB+XGB+HGB+LGBM+AB)+LSTM provided the strongest stacking compromise, with recall of 0.7255, G-mean of 0.8254, and ROC-AUC of 0.9270.SHAP analyses of HGB with SVM-SMOTE and GRU with SMOTE-ENN identified debt ratio, profitability, retained earnings, solvency, and asset turnover indicators as influential predictors.
4 Discussions
The study shows that bankruptcy-model performance depends strongly on minority-class detection and resampling strategy, with deep models and SMOTE-ENN generally favoring recall and balanced performance. Stacking offered a useful compromise, while SHAP highlighted economically interpretable drivers of bankruptcy risk.
- Resampling effects: SVM-SMOTE and SMOTE-Tomek favored accuracy and specificity, whereas SMOTE-ENN produced stronger recall and average G-mean across stacked models.The study attributes this behavior to differences in the class structure induced by hybrid resampling.
- Standalone machine-learning models: HGB was the strongest machine-learning classifier by ROC-AUC across all three resampling strategies, while LGBM achieved the highest accuracy and specificity.This contrast suggests different boosting models emphasized discrimination versus conservative bankrupt-class assignment.
- Standalone deep-learning models: Under SMOTE-ENN, GRU achieved ROC-AUC of 0.9431, while LSTM and MLP achieved recall of 0.8824.Neural models generally exceeded machine-learning models on recall and often on G-mean under the tested resampling strategies.
- Stacking ensembles: SMOTE-ENN with (GB+XGB+HGB+LGBM+AB)+LSTM achieved recall of 0.7255, G-mean of 0.8254, and ROC-AUC of 0.9270 among stacking ensembles.The stacking result provided a compromise between sensitivity and specificity but did not automatically exceed the strongest standalone deep models.
- Feature interpretation: Bankruptcy risk was primarily associated with retained earnings to total assets, ROA, net worth to assets, debt ratio, total asset turnover, persistent EPS, and operating expense rate.These predictors span profitability, solvency, capital structure, and operational efficiency.
5 Conclusion & Recommendations
The study concludes that balanced bankruptcy prediction should prioritize minority-class detection, interpretability, and operational relevance rather than accuracy alone. It also identifies structured-data, single-country evaluation as a boundary and recommends broader, longitudinal, multimodal validation.
- Conclusions: Class imbalance shapes model rankings, error distributions, and practical usefulness in bankruptcy prediction.The study therefore frames evaluation around discrimination, interpretability, and operational relevance rather than accuracy alone.
- Conclusions: The framework combines consensus feature selection, hybrid resampling, heterogeneous learning architectures, and SHAP-based interpretation.The study presents this combination as advancing a more balanced view of bankruptcy-model performance.
- Recommendations: Structured financial-ratio data alone still support useful bankruptcy prediction, although recent research reports incremental gains from textual and other nonfinancial inputs.The paper positions its results as complementary to, rather than opposed to, multimodal bankruptcy research.
- Limitations and recommendations: The analysis used a single Taiwanese bankruptcy dataset and mainly structured financial ratios, limiting generalizability across countries, industries, and regulatory environments.The authors recommend larger multi-country datasets, longitudinal and out-of-time testing, and additional financial, market, and textual information.