Cost-sensitive stacking ensemble learning for company financial distress prediction

Abstract

Financial distress prediction (FDP) is a topic that has received wide attention in the finance sector and data mining field. Applications of combining cost-sensitive learning with classification models to address the FDP problem have been intensely attracted. However, few combined cost-sensitive learning and Stacking to predict financial distress. In this article, a cost-sensitive learning method for FDP, namely cost-sensitive stacking (CSStacking), is put forward. In this work, a two-phase feature selection method is used to select the optimal feature subset. A CSStacking ensemble model is developed with selected features to make a final prediction. The paired T test and non-parametric Wilcoxon test are employed to check the significant differences between CSStacking and benchmark models. An experiment over Chinese listed company dataset is designed to investigate the effectiveness of CSStacking. The experimental results prove that CSStacking can forecast listed companies’ financial distress five years ahead and improves the identification rate of financially distressed companies, highlighting its potential to reduce economic losses caused by misclassifying financially distressed companies. The results of comparing CSStacking with four types of benchmark models show that CSStacking performs significantly better than benchmark models. Furthermore, the findings illustrate that “asset-liability ratio”, “current ratio”, “quick ratio”, and “industry prosperity index” are critical variables in predicting financial distress for Chinese listed companies.