To boost or not to boost? XGBoost and DCC-GARCH integration for mean-variance portfolio optimization
Abstract
This study examines whether integrating machine learning-based return forecasting and dynamic covariance estimation into a mean-variance portfolio framework produces measurable improvements over a conventional benchmark. Three strategies are constructed and evaluated over a five-year out-of-sample window from January 2021 to December 2025, applied to a universe of ten global exchange-traded funds spanning US fixed income, developed equity markets, and emerging equity markets. The benchmark strategy (BM) forms portfolios using historical sample means and sample covariance. The HR+DCC strategy replaces the sample covariance with a DCC-GARCH covariance forecast while retaining historical return estimates. The XGB+DCC strategy further replaces the return input with a 20-trading-day-ahead log return forecast from a hyperparameter-tuned XGBoost model trained on technical and macroeconomic features, with dimensionality reduced through principal component analysis. All three strategies maximize the Sharpe ratio subject to long-only constraints with a 30% per-asset weight cap and are rebalanced monthly using a rolling three-year estimation window. Before transaction costs, XGB+DCC achieves the highest cumulative log return (38.14%) and annualized Sharpe ratio (0.416), compared with 0.377 for BM and 0.353 for HR+DCC, indicating that the XGBoost return forecast adds information value in gross terms. However, XGB+DCC exhibits substantially higher monthly turnover (45.77% versus 10.07% for BM and 11.88% for HR+DCC), and once a 0.1% proportional transaction cost is applied, its net Sharpe ratio (0.343) converges toward that of BM (0.358), with a breakeven cost of only 0.077%. Jensen alpha estimates for both model-based strategies are statistically insignificant relative to the benchmark. These findings suggest that the practical value of machine learning-based return forecasting in this setting depends critically on implementation costs, and that turnover-penalizing construction or lower-frequency rebalancing may be necessary for gross gains to survive in net terms.