Abstract Corporate bankruptcy prediction plays a critical role in financial risk management, as it enables the early identification of financially distressed firms and helps reduce potential economic losses. However, most bankruptcy datasets are highly imbalanced and contain a large number of features, making it difficult for traditional machine learning models to accurately detect bankrupt firms. To address these challenges, this study proposes a hybrid framework that integrates Generative Adversarial Networks (GANs) for minority-class data augmentation, Recursive Feature Elimination (RFE) for feature selection, and a Random Forest classifier for prediction. The proposed method is evaluated using the Taiwan bankruptcy dataset, which consists of 6,819 companies described by 95 financial ratios, with only 2.6% of firms classified as bankrupt. The results suggest that the proposed approach improves minority-class detection performance and enhances the identification of bankrupt firms. Experimental evaluation using stratified 5-fold cross-validation demonstrates that the proposed framework outperforms baseline models and traditional resampling techniques such as SMOTE and ADASYN in terms of minority-class recall and PR-AUC. The results confirm that the integration of GAN-based augmentation, feature selection, and threshold optimization provides significant improvements for highly imbalanced bankruptcy prediction tasks.
Mundkar et al. (Tue,) studied this question.