Author(s)

Emma Li, He YANG, Jiapei Li, Yifang Cai

Corresponding Author

Jiapei Li

Abstract

Early warning of financial distress is of great significance to the healthy development of enterprises and stakeholders’ decision-making. Nowadays, the data-driven machine learning method has replaced the statistical methods to become the mainstream method of financial early warning. However, the complex model brings higher prediction accuracy, and brings the model unexplained embarrassment as well. In this paper, the ensemble learning algorithm GBDT(Gradient Boosting Decision Tree) is introduced to establish a financial distress early warning model, and the interpretability of the model is studied based on the SHAP(SHAPley Additive exPlanations) framework. Then the empirical study on more than 2000 listed manufacturing companies in China shows that the GBDT model has better generalization ability than the model based on Logistic regression, and SHAP gives a clear explanation about how financial distress contributed by feature variables. This study is valuable not only in theory but also in application for financial distress risk early warning.

Keywords

Financial Distress, Early-warning, GBDT, SHAP