A dynamic asset allocation approach under technical analysis and machine learning classification

This research evaluates the effectiveness of machine learning models by comparing the performance of seven models in predicting the return movement of four asset types over different return periods. The assets include S&P 500 Futures, Gold Futures, US 10Y Treasury Bond Futures, and a risk-free investment. The study considers 1-week, 4-week, and 12-week forward return periods. The analysis assesses model accuracy and portfolio performance using data spanning two decades. ML-based portfolios modestly outperform an equal-weighted benchmark in terms of risk-adjusted metrics over longer horizons, even though the differences are generally insignificant. However, the study emphasizes the importance of considering diverse economic cycles to understand asset behaviour comprehensively. Nevertheless, the study did not find significant evidence that combining machine learning prediction with dynamic asset allocation can outperform passive portfolio investment. The study underscores the potential of non-linear models in short-term trading but also notes limitations in adaptiveness, threshold design, and model generalization under atypical conditions.