Abstract：

This paper uses a sample of non-financial A-share listed companies in China from 2013 to 2022. First, we perform SMOTE oversampling on the data to increase the number of minority class samples and achieve class balance. Second, we employ machine learning methods to predict and analyze corporate debt risk. Empirical results show that the prediction performance of two ensemble learning models, random forest and XGBoost, is significantly superior to that of the traditional logit model. Furthermore, the prediction accuracy does not necessarily improve significantly with larger datasets. Furthermore, the factors influencing corporate debt risk vary from year to year. This study's conclusions provide investors, corporate managers, and other stakeholders with a more accurate and effective approach to corporate debt risk assessment and can also inform future corporate debt risk prevention strategies.

Key Words：

machine learning; SMOTE oversampling processing; debt risk; capital liability ratiocapital-to-debt ratio