A machine learning framework for multi-market portfolio optimization: Evidence from U.S. stocks and cryptocurrencies
TL;DR
The results show that the predictive models provide useful signals for portfolio construction, with all optimized strategies achieving markedly lower risk than the naïve portfolio for the same expected return level.
Abstract
This study presents an integrated framework for multi-market portfolio optimization that integrates machine-learning-based return forecasting with classical and downside-oriented risk models. Using daily data for Bitcoin, Ethereum, BNB, Microsoft, and Tesla, the XGBoost algorithm is employed to predict short-term returns, after which the forecasts are used to construct optimal portfolios under the Mean-Variance (MV), Mean-Semi-Variance (MSV), and Mean-Absolute Deviation (MAD) models. A naïve equal-weighted portfolio is included as a benchmark. The results show that the predictive models provide useful signals for portfolio construction, with all optimized strategies achieving markedly lower risk than the naïve portfolio for the same expected return level. Tangency portfolios, particularly the Maximum Sortino and Maximum Sharpe-AD configurations, delivered the highest efficiency gains and the strongest performance on out-of-sample data. Overall, the findings demonstrate that combining machine learning with risk-sensitive optimization tools can improve decision-making in markets characterized by volatility and mixed asset dynamics.