This study proposes a Genetic Algorithm–optimized LightGBM (GA – LightGBM) model for forecasting stock volatility of ASM.VN, demonstrating superior performance compared with traditional models (GARCH, SVMRegressor, XGBoost) and deep learning architectures (GRU, TabNet, Transformer, CNN, LSTM). The model achieves MAE = 0.0470; RMSE = 0.0731; R² = 0.8480; STD = 0.1480, with a mean R² = 0.805 across 5-fold cross-validation, indicating high accuracy and stability. Predicted volatility over 06 February 2025 – 06 March 2025 ranges from 0.1419 to 0.2154, reflecting moderate adjustment risk. SHAP analysis identifies key original features in descending order: HV, BB – width, BB – lower, VWAP, ATR, MA_21, BB – upper, OBV, highlighting the roles of market fluctuations, price amplitude, liquidity, and medium-term trends in shaping forecasts. Sensitivity analysis indicates colsample – bytree and num – leaves exert the strongest influence (ΔR² = 2.02% and 0.81%, respectively), whereas other hyperparameters produce minor variations (<0.6%), confirming model robustness. During the period starting 09 April 2025, US retaliatory tariffs (+46%) triggered strong volatility spikes in energy stocks, particularly ASM.VN (+130%), HDG.VN (+93%), BCG.VN (+60%), with differentiated forecasts persisting through September 2025. The GA – LightGBM model demonstrates strong potential for guiding investment strategies, risk management, and policy-sensitive market analysis. While empirical results are robust, economic implications emphasize practical benefits including risk-adjusted returns, improved hedging efficiency, and reduced pricing errors in derivative markets. Caution is warranted under “black swan” events or structural market changes, necessitating periodic retraining and hyperparameter monitoring to maintain predictive reliability.
