Do Fusion Models Improve Volatility Forecasting? Evidence from Developed, Emerging, and Frontier Stock Markets

This paper investigates the efficacy of fusion models, which integrate machine learning (ML) techniques and traditional time series volatility models, in forecasting stock market volatility across developed, emerging, and frontier markets. Using comprehensive data from the United States, United Kingdom, China, India, and frontier markets, we benchmark fusion models combining GARCH variants with ensemble algorithms (RandomForest, XGBoost, CatBoost) against standalone ML and traditional time series models (GARCH, EGARCH, GJR-GARCH). Our empirical results show that fusion models consistently outperform standalone models in predictive accuracy, measured by reduced root mean square error (RMSE), mean absolute error (MAE) and Coefficient of Determination (R2). Specifically, frontier markets demonstrate the most significant improvements, indicating that fusion modeling is particularly beneficial where volatility signals are weaker and markets less efficient. Furthermore, trading simulations based on these predictions highlight that fusion models not only achieve superior volatility forecasts but also provide better trading outcomes, reflected in increased profitability, enhanced annualized returns, and reduced drawdowns. These findings suggest that integrating ensemble ML methods with volatility modeling significantly enhances forecasting performance and practical utility across diverse financial markets.