XAI-Driven Bankruptcy Prediction Using Hybrid Ensembles

Predicting corporate bankruptcy is a critical task in finance, often complicated by severely imbalanced datasets where bankruptcy cases are rare. This research introduces a comprehensive framework designed to enhance the accuracy and interpretability of bankruptcy predictions, particularly for the minority class. The process begins with consensus-based feature selection, reducing the input to 23 robust financial variables from the Taiwanese Bankruptcy Prediction dataset. To address data imbalance, the study employs hybrid resampling techniques like SVM-SMOTE, SMOTE-Tomek, and SMOTE-ENN. Various machine learning and deep learning models are then compared, including gradient boosting, XGBoost, LightGBM, AdaBoost, RNN, LSTM, GRU, DNN, and MLP. The framework further explores hybrid stacking ensembles, combining machine learning classifiers as base learners with deep learning models as meta-learners. Performance is evaluated using multiple metrics, with SMOTE-ENN showing superior minority-class detection. The GRU model with SMOTE-ENN achieved the best standalone balance, while a specific stacking ensemble (SMOTE-ENN with GB+XGB+HGB+LGBM+AB)+LSTM provided the strongest compromise between sensitivity and specificity. Crucially, SHAP analysis was used to explain feature contributions, identifying leverage, profitability, solvency, and operational efficiency as the most influential predictors, offering valuable insights for early warning systems.