Combining multiple data resampling methods and classifier ensembles for better financial distress prediction: homogeneous and heterogeneous approaches

Financial distress prediction (FDP) is a critical task for financial institutions and is typically framed as a class imbalance learning problem. To address this challenge, this paper proposes two ensemble-based strategies: the homogeneous and heterogeneous approaches, which combine multiple data re-sampling algorithms to generate diverse re-balanced training sets for classifier construction. Experimental results on seven FDP datasets demonstrate that the heterogeneous approach, which integrates under-, over-, and hybrid sampling methods with their optimal imbalance ratio settings, achieves superior performance in terms of AUC, particularly when applied with the LightGBM and XGBoost classifiers. Regarding Type I error, the heterogeneous combinations consistently outperform the homogeneous and other baseline approaches across various classifiers. The generalizability of the proposed methods is further validated using 37 additional class-imbalanced datasets from different domains, where the heterogeneous approach again shows the most robust performance. These findings suggest that the proposed models can serve as effective decision support tools for financial institutions to enhance credit risk evaluation and lending strategies. From a policy perspective, adopting such predictive frameworks can improve financial stability by reducing exposure to high-risk loans and enabling more accurate early warning systems for economic distress.