Evaluating machine learning-based portfolio strategies versus modern portfolio theory in robo-advisors

Abstract

The robo-advisory market is projected to grow nearly tenfold by 2034. Despite its well-established shortcomings in the current unstable market, most automated asset management systems continue to rely on Modern Portfolio Theory. The emergence of machine learning raises the question of whether it can replace Modern Portfolio Theory as the primary asset allocation engine in robo-advisors. This study evaluates whether machine learning-driven portfolio strategies can outperform a traditional Modern Portfolio Theory benchmark in a robo-advisor context. Empirical analysis is conducted on daily data from twenty US-listed assets covering the period from February 2021 to February 2026, employing Random Forest and Naïve Bayes classifiers, trained asset-by-asset with a three-class 21-day directional target. Predictions are translated into portfolio weights via score-weighted aggregation and rebalanced monthly, then benchmarked against a Global Minimum Variance portfolio using risk-adjusted metrics. Both classifiers struggled to achieve balanced classification skill. Performance ranged from broadly accurate on ETFs to worse than random chance on high-volatility stocks. While the ML strategies generated comparable cumulative returns to MPT (RF: 19.12%, NB: 20.32%, MPT: 21.06%), they do so at double the volatility (~ 19.6% vs 10.15%) and double the maximum drawdown (~ −20% vs −9.4%). Sharpe ratios of 0.76 and 0.80 against MPT's 1.61 confirm ML did not outperform on a risk-adjusted basis. Notably, Naïve Bayes matches Random Forest despite weaker classification skill, suggesting that weight allocation matters as much as predictive accuracy. Within the conditions of this study, classification-driven ML cannot replace MPT in robo-advisors. The findings position ML as a complementary overlay, with ETF-dominated universes the most appropriate deployment scope.