Generative deep learning for class imbalance in credit risk modeling: comparative evaluation of synthetic oversampling methods and an elastic oversampling extension

This paper applies generative deep learning–GAN, VAE, and CVAE–to oversample rare defaults in credit risk modeling, with a posterior binarization procedure introduced to preserve the marginal distribution of dummy features. Using firm- and individual-level default datasets, we compare these methods against SMOTE, SMOTE-Tomek, and ADASYN under multiple classifiers. Empirical results show that synthetic oversampling improves recall and disperses feature importance, while generative oversampling tends to sharpen class boundaries, favor precision, and concentrate feature importance. To balance these complementary strengths, we introduce Elastic Oversampling, a parallel hybrid of generative and synthetic approaches with a tunable mixing ratio for the precision–recall trade-off. Elastic Oversampling emphasizes distinctive minority cases through generative methods while constraining unrealistic samples with synthetic methods, offering a flexible and policy-sensitive framework for class-imbalanced credit risk prediction.