Cost-sensitive ensemble learning for bankruptcy prediction under extreme class imbalance

Abstract Corporate bankruptcy prediction is a high-stakes artificial intelligence (AI) task characterized by extreme class imbalance and asymmetric misclassification costs. Although ensemble learning models have shown strong predictive performance, most existing studies rely on cost-insensitive metrics and fixed decision thresholds, which can misrepresent real-world decision utility. This study reframes bankruptcy prediction as a decision-centric AI problem and proposes a cost-sensitive ensemble learning framework that explicitly decouples probabilistic prediction from decision-making. Ensemble models are trained to produce calibrated risk estimates and are subsequently combined with cost-aware threshold optimization to minimize expected misclassification cost. An extensive evaluation on a real-world bankruptcy dataset using repeated, leakage-free cross-validation integrates imbalance-aware and decision-centric metrics, including PR-AUC, F-score, expected cost, and calibration measures. The results show that model rankings change substantially under decision-centric evaluation. Boosting-based ensembles provide the most favorable balance between minority detection, probability reliability, and decision cost, while data-level resampling via SMOTEENN yields limited benefits once cost-sensitive optimization is applied. Overall, the study highlights the importance of separating prediction from decision-making in imbalanced AI systems and offers practical guidance for deploying ensemble models in high-stakes risk assessment.