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

When firms teeter on the edge of insolvency, early warning signals can make the difference between a rescue plan and a costly liquidation. Yet most public datasets contain far more healthy companies than bankrupt ones, squeezing the signal from the minority class into noise. A recent pre‑print tackles this imbalance by marrying several resampling tricks with a stacking ensemble and then opening the black box with explainable AI techniques.

What You Need to Know

The authors start with the Taiwanese Bankruptcy Prediction set from the UCI repository. Five different feature‑selection algorithms are run, and a consensus rule keeps only the variables that appear in at least three of the five lists, shrinking the input space to 23 robust predictors.

To counter the extreme class imbalance, three hybrid resampling strategies are applied to the training data: SVM‑SMOTE, SMOTE‑Tomek, and SMOTE‑ENN. Each method creates a balanced version of the data, allowing the models to see enough minority‑class examples without overwhelming the majority class.

Five gradient‑based learners—gradient boosting, XGBoost, HistGradientBoosting, LightGBM, and AdaBoost—are trained on each resampled version. Their predictions are then combined in a stacking meta‑learner, which learns how to weight the base models for optimal bankruptcy detection.

Why It Matters

Regulators, creditors, and auditors need reliable early‑warning tools to intervene before a firm collapses. Improving the recall of bankruptcies while keeping false alarms low can save millions in potential losses and protect stakeholders.

By layering explainable AI—such as SHAP values and feature importance—onto the stacked model, the study moves beyond pure accuracy. Stakeholders can see which financial ratios drive a high‑risk score, turning a opaque prediction into actionable insight for monitoring and decision‑making.

Key Details

• Consensus feature selection retained 23 variables from an original set of 95.
• Three resampling methods (SVM‑SMOTE, SMOTE‑Tomek, SMOTE‑ENN) were each applied to the training split.
• Five base classifiers were gradient boosting, XGBoost, HistGradientBoosting, LightGBM, and AdaBoost.
• A stacking meta‑learner combined the base models’ outputs using logistic regression.
• Explainability was achieved with SHAP dependence plots and global feature importance scores.

What’s Next

Future work could test the framework on other international bankruptcy datasets and explore cost‑sensitive learning to further tune the trade‑off between precision and recall. Additionally, integrating temporal features—such as quarterly trend changes—might improve the model’s ability to signal distress earlier in a firm’s lifecycle.

📌 Source: Arxiv Ml