Heterogeneous performance of machine learning models in financial market forecasting

Forecasting financial markets remains challenging due to nonlinear dynamics, structural changes, volatility clustering, and uncertainty transmission across asset classes. Although machine learning methods have improved predictive modeling, limited evidence exists regarding their relative performance across heterogeneous financial markets under consistent evaluation conditions. This study addresses this gap by comparing five widely used machine learning approaches: Ridge Regression (Ridge), Least Absolute Shrinkage and Selection Operator Regression (Lasso), Random Forest Regression (RF), Bootstrap Aggregation (Bagging), and Gradient Boosting Machine (GBM). The analysis focuses on forecasting the Volatility Index (VIX), S&P 500 index, gold prices, Natural Gas Future (NG-F), and Brent Crude Oil Future (Brent-F) using monthly data from January 2000 to July 2025. All models are evaluated within a unified forecasting framework using Mean Absolute Error (MAE), Mean Squared Error (MSE), and the coefficient of determination (R²). The results demonstrate substantial heterogeneity in model performance across financial markets. Ridge and Lasso achieve strong predictive accuracy in markets characterized by stable relationships and high correlation among predictors, while ensemble methods, particularly RF and GBM, provide greater flexibility in capturing nonlinear interactions and complex market dynamics. The findings indicate that no single algorithm consistently dominates across all financial indicators; instead, forecasting performance depends on the interaction between model characteristics and the economic structure of each market. This study highlights the importance of context-specific model selection and provides practical insights for investors, policymakers, and risk managers seeking robust forecasting approaches under changing financial conditions.