Stock Return Prediction Using Macroeconomic Drivers: The Case of the KOSPI Index

This study evaluates the predictive performance and investment value of various models for forecasting monthly KOSPI returns using macroeconomic and financial variables. Our empirical findings show that the rolling-window LASSO and Random Forest models significantly outperform other competing approaches, including standard linear regression and deep learning methods. Using three different hyperparameter tuning criteria, we find that while the performance of the Random Forest model is highly sensitive to hyperparameter choices, the rolling-window LASSO model, which accounts for time-varying relationships between KOSPI returns and predictive variables, consistently delivers superior predictive accuracy and investment performance. Furthermore, no single hyperparameter tuning criterion consistently yields optimal investment outcomes, underscoring the importance of employing multiple evaluation metrics for hyperparameter tuning in practical applications.