Class-imbalanced dynamic financial distress prediction based on random forest from the perspective of concept drift

An effective enterprise financial distress prediction system is one of the key measures to prevent and resolve enterprise debt risk. However, it lacks enough and deep research on how to dynamically construct ensemble models for class-imbalanced financial distress prediction under the situation of concept drift. From the perspective of concept drift, this paper constructs a random forest model for dynamic prediction of corporate financial distress by considering class imbalance between financially distressed and non-distressed enterprises, to improve the performance of dynamic financial distress prediction. Using the sample data of the public companies listed in the Shanghai and Shenzhen Stock Exchange of China from 2010 to 2020, this paper carries out the empirical research and finds that there exists financial distress concept drift for Chinese listed companies. The full-memory rolling time window mechanism and the fixed-width rolling time window mechanism can improve the prediction effect of models and the fixed-width rolling time window mechanism is better than the full-memory rolling time window mechanism. The combination of SMOTE and random under-sampling can solve the class-imbalance problem to some extent. The analysis of the importance of indicators shows that financial indicators of profitability are more informative for Chinese listed companies’ financial distress prediction. In addition, dynamic financial distress prediction model based on random forest integrated with resampling mechanism significantly outperforms other models.