Forecasting the default risk of Chinese listed companies using a gradient-boosted decision tree based on the undersampling technique

Shanshan Wang, Guotai Chi, Ying Zhou and Li Chen

Abstract

Default prediction is of interest to the creditors, customers and suppliers of any firm as well as to policymakers and current and potential investors. Imbalanced classification for default prediction is considered a crucial issue. Therefore, this study proposes a default risk prediction model using a gradient-boosted decision tree (GBDT) based on the random undersampling (RUS) technique. We build a default prediction model based on 29 indicators and five different time windows. The model has two steps. First, the proposed RUS-GBDT model adopts the undersampling approach to generate different training samples based on the imbalance ratio of the training data. Then, the parameter of the GBDT is adaptively tuned with the area under the receiver operating characteristic curve of the predictive model for the selected training sample. We analyze the optimal imbalance ratio of the different training samples and compare the model’s prediction performance with that of several other classification methods including logistic regression and support vector machines. Our experimental results demonstrate that the proposed model performs better than the other classifiers with respect to predicting and classifying the default status of listed companies in China.