Global stock market forecasting: Insights from series and parallel combination of machine learning models

In the field of stock market forecasting, a meticulously devised combination of machine learning models to improve forecast accuracy and investment returns is of considerable value. This research focuses on integrating Lasso, Multilayer Perceptron (MLP), and XGBoost models using both series and parallel configurations to enhance the predictive accuracy of global stock index returns across nine major markets. The serial integrations include the progression from Lasso to MLP to XGBoost (Lasso-MLP-XGBoost-Series), while the parallel configuration specifically refers to the Lasso-MLP-XGBoost-Parallel model. The results demonstrate that both the series and parallel models generally outperform individual models and the traditional Buy-and-Hold (B&H) strategy, with the Lasso-XGBoost and Lasso-MLP-XGBoost-Parallel models showing strong return and risk-adjusted performance, particularly in developed markets. Additionally, a detailed analysis of the Variable Importance in the Projection (VIP) and SHapley Additive exPlanations (SHAP) values indicates that different lags of the VIX index are critical for forecasting in different regions, and that return transmission across markets within short time windows reflects the dominant role of the U.S. dollar in global equity dynamics.