Understanding Corporate Bond Defaults in Korea Using Machine Learning Models^*

Abstract

We investigate corporate bond defaults from 1995 to 2020 using hand‐collected data from hard‐copy publications in Korea. Using an under‐sampling method, we construct default prediction models based on machine learning models as well as a logistic model. The empirical results show that the random forest model outperforms the others. However, regardless of the models used, model performance in financial crisis periods is significantly worse than it is in non‐crisis periods. This finding suggests the need for additional information to improve model performance during crises when the default prediction is the most relevant. Furthermore, the dominant predictor of defaults before the global financial crisis was the debt ratio, while subsequently, the coverage ratio has become the most important predictor.