EARLY PREDICTION OF FINANCIAL DISTRESS RISK IN VIETNAMESE LISTED INDUSTRIAL FIRMS USING MACHINE LEARNING

Abstract

Early prediction of financial distress risk is important for providing timely warning signals to investors, creditors, and regulators. This study examines 91 Vietnamese listed industrial firms over the period 2015–2024 and compares a traditional Logistic Regression model with three machine learning algorithms: Random Forest (RF), Extreme Gradient Boosting (XGBoost), and K-Nearest Neighbors (KNN). The dependent variable captures a firm’s financial distress status in the following year, while the explanatory variables include conventional financial ratios, firm size, and lagged GDP growth. The empirical results for the test period show that all machine learning models outperform Logistic Regression in terms of overall predictive performance. RF achieves the highest Area Under the ROC Curve (AUC = 0.8741), followed by XGBoost (0.8578) and KNN (0.8030), while Logistic Regression records the lowest value (0.7232). At the 0.2 probability threshold, XGBoost provides the strongest distress detection, with the highest Recall of 76.92% and the highest F2-score of 71.68%, whereas RF offers a better balance between detecting distressed firms and controlling false alarms. Additional sensitivity analysis at the 0.3 probability threshold confirms the stability of the main findings. Furthermore, SHAP-based feature importance indicates that asset structure, profitability, leverage, and cash-flow capacity are the most influential predictors in the tree-based models. These findings support the use of machine learning, particularly RF and XGBoost, for early warning systems in Vietnamese listed industrial firms.