This study investigates whether machine learning effectively processes high-dimensional data, a challenging task for traditional predictive models, to optimize portfolio strategies. Using daily data from December 2004 to July 2024, we compare various machine-learning models for asset allocation in an all-weather portfolio comprising exchange-traded funds for the S&P 500, long-term Treasury bonds, and gold. We find that the LASSO and elastic net models exhibit superior overall performance, whereas tree-based models excel in forecasting long-term Treasury bond returns. Portfolio strategies employing these models achieve Sharpe ratios near 0.70, substantially outperforming static benchmarks. The results demonstrate that machine learning can optimize portfolio performance in practical investment settings.
Ha et al. (Mon,) studied this question.