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

Resampling strategy significantly impacts model performance: The study found that different data balancing techniques (SVM-SMOTE, SMOTE-Tomek, and SMOTE-ENN) materially shaped how well models performed, with SMOTE-ENN delivering the strongest detection of bankrupt companies—a critical outcome for financial risk management.

Explainability is essential for financial AI systems: The researchers used SHAP (a specific explainable AI technique) to identify which financial features most influenced bankruptcy predictions, demonstrating that understanding why an AI system makes decisions is as important as the accuracy of those decisions in high-stakes financial contexts.

Hybrid ensemble approaches outperform single models: By combining multiple machine learning and deep learning models together (stacking ensembles), the framework achieved better balanced performance than any single model alone, suggesting that responsible AI deployment benefits from diverse algorithmic perspectives rather than relying on one approach.

Minority-class detection requires specialized techniques: The study addressed a real-world challenge where bankrupt companies are rare in datasets, requiring specialized resampling and ensemble methods to ensure AI systems don't miss actual bankruptcies—a critical consideration when deploying AI for financial risk assessment.

Feature selection reduces complexity without sacrificing performance: By using consensus-based methods to reduce 95+ financial variables down to 23 robust ones, the framework created a more interpretable and maintainable system, supporting the principle that simpler, explainable AI systems are preferable when they perform comparably to complex alternatives.