Predicting financial distress using machine learning approaches: Evidence China

Abstract

This study uses machine learning techniques to construct financial distress prediction (FDP) models for Chinese A-listed construction companies and compares their classification performance with conventional Z-Score models. Three machine learning algorithms (Classification and Regression Tree, AdaBoost, and CUSBoost) are used to generate machine-learning-based classifiers, and four Z-Score models (Altman Z-Score, Sorins/Voronova Z-Score, Springate, and Z-Score of Ng et al.) are selected for comparison. The sample comprises 1782 firm-year observations from Chinese A-listed construction companies on the Shenzhen and Shanghai Stock Exchanges from 2012 to 2021. The out-of-sample predicting performance of the classifiers are measured using the areas under the receiver operating characteristic curve (AUC) and under the precision-recall curve (AUPR). In additional tests, Pearson's correlation coefficients and the variance inflation factor are utilized to identify correlations among the raw financial predictors, while principal component analysis is used to address high-correlation issues among the features. Results confirm that machine learning classifiers can effectively predict financial distress for Chinese A-listed construction companies and are more accurate than Z-Score models. Furthermore, the CUSBoost classifier is identified as the most precise model based on the AUC and AUPR metrics in both primary and additional tests. This study addresses the gap concerning the application of machine learning in FDP for Chinese-listed construction companies. Additionally, the CUSBoost Algorithm is introduced into the field of FDP research for the first time. Through the comparison of machine learning and Z-Score models, this study also contributes to the literature related to the contrast between machine learning and statistical modeling techniques.