Explainable machine learning for corporate financial distress prediction in Vietnam

Abstract

Corporate financial distress generates substantial costs for firms and the broader economy, making timely early-warning tools especially important in emerging markets. This study develops an interpretable distress-prediction framework for Vietnamese listed companies by assessing whether modern classifiers outperform a theory-driven logistic benchmark and by identifying economically meaningful risk drivers. Using firm-year data from 2014 to 2024, the analysis combines accounting indicators, market-based signals, and a macroeconomic factor. Models are trained on 2014 to 2021 with time-aware cross-validation and evaluated on 2022 to 2024 using ranking metrics, classification performance, calibration, and Shapley additive explanations. The comparison includes logistic regression, decision tree, random forest, support vector machine with radial basis function kernel, multilayer perceptron, and extreme gradient boosting. Ensemble models perform best: random forest achieves an area under the receiver operating characteristic curve of about 0.946 and a balanced F1 score near 0.74, closely followed by extreme gradient boosting. Explanations show that current operating profitability is the strongest nonlinear driver, with distress risk rising sharply once firms become loss-making. Accumulated profitability, working-capital liquidity, firm size, market valuation, and inflation add further predictive value. The study offers an auditable framework for investors, lenders, managers, and supervisors in Vietnam.