PBESProceedings of Business and Economic Studies2209-26412209-265XBio-Byword Scientific Publishing Pty. Ltd.10.26689/PBES.v9i9.15611ArticleModel Selection and Threshold Selection in Financial Distress Prediction: An Empirical Analysis Based on Listed Manufacturing Companies in Chinahttps://artdesignp.com/journal/PBES/9/9/10.26689/PBES.v9i9.15611DuanSiqi2026992026-09-30This paper analyses 3,837 manufacturing companies listed on the Shanghai and Shenzhen A-share markets between 2010 and 2024. It uses two consecutive years of losses as an indicator of financial distress and employs the T−2 offset forecasting method to avoid forward-looking bias. The data are divided chronologically into training and testing sets, and a comprehensive comparison is made of five commonly used early warning methods. Furthermore, the paper proposes a PAT dynamic threshold adjustment approach that accounts for time-series characteristics. The study concludes that there is limited scope for improvement in machine learning models; XGBoost only marginally improves the AUC relative to the baseline Logit model. If 0.5 is used as a fixed threshold, the early warnings issued by the model are largely ineffective, and non-threshold models are of little practical use. The threshold is the key determinant of the success of early warning; it is a more powerful indicator of cost issues than the model itself. Dynamic PAT thresholds can significantly reduce the cost of early warnings and markedly improve the identification of financial distress. From a practical perspective, financial risk early warning systems should focus on cost-oriented threshold setting rather than simply increasing model complexity.Financial distress prediction, Decision threshold, Model selection, Class imbalance, PAT dynamic threshold

[1] Li C, 2018, Construction of a Financial Risk Early Warning Model for Enterprises Based on Logistic Regression. Statistics and Decision-Making, 34(6): 185–188. [2] Guo E, 2020, Prediction of Financial Distress in Chinese A-share Listed Companies Based on the Logistic Model. Business Accounting, (10): 35–39. [3] Wang R, 2019, Prediction of Financial Distress in Listed Companies Based on Lasso-Logistic and XGBoost, thesis, Zhongnan University of Economics and Law. [4] Xiang Y, 2022, A Study on the Early Warning of Financial Distress in Listed Companies Based on Machine Learning, thesis, Southwest University. [5] Song Y, 2024, Prediction of Financial Distress in Information Technology Enterprises Based on Random Forest and SVM Algorithms. Modern Computers, 30(9): 43–50. [6] Zhang X, 2023, A Study on the Early Warning of Financial Distress in Listed Companies Based on Lasso-GWO-SVM, thesis, Shanghai Normal University. [7] Zou F, 2023, Prediction of Financial Distress in Small and Medium-sized Enterprises Based on Machine Learning. Information Systems Engineering, (12): 68–71.