DETECTION OF EARNINGS MANAGEMENT USING MACHINE LEARNING: INSIGHTS FROM VIETNAM

This study investigates the application of machine learning algorithms to detect earnings management among non-financial firms listed on the Vietnamese stock market (HOSE and HNX) from 2000 to 2024, results to dataset of 8,371 observation from 630 listed firms. Using discretionary accruals estimated through the Modified Jones Model with performance adjustment as the proxy for earnings management, firm-year observations are classified into managing and non-managing groups. A set of financial ratios, firm characteristics, and governance indicators are employed as predictors. Five machine learning algorithms—logistic regression, support vector machine, random forest, multilayer perceptron, and extreme gradient boosting (XGBoost) are trained and evaluated using 10-fold cross-validation. Among these, XGBoost consistently outperforms others across key metrics, confirming its superior predictive performance and reliability. SHAP analysis further reveals the relative importance of key variables such as firm size, accrual quality, and liquidity indicators. The study contributes methodologically by demonstrating the power of advanced classification models in detecting earnings manipulation, and contextually by offering empirical insights specific to an emerging market. These findings underscore the potential of machine learning as a decision-support tool for regulators, auditors, and investors concerned with financial reporting quality.