Bankruptcy prediction using synthetic sampling

Abstract

The prediction of corporate bankruptcy has been widely studied. However, low bankruptcy rates lead to highly imbalanced bankruptcy class distributions, increasing the difficulty of accurately predicting a firm’s bankruptcy. Based on a large sample of 1824 U.S. firms, the study shows that classification accuracy significantly improves when the training dataset is balanced using the synthetic minority oversampling technique or one of its extensions. The results indicate that combining SMOTE with cluster-based undersampling leads to the best classification performance, and the increase in accuracy, specifically in terms of recall and AUC, is significant, thus justifying synthetic sampling when training a bankruptcy prediction classifier.