Investing in AI Stocks: Machine-Learning Forecasts, Portfolio Strategies, and Stress Tests

This paper examines whether machine-learning forecasts of artificial intelligence (AI) exchange-traded fund (ETF) prices can improve portfolio diversification, reduce downside risk, and increase risk-adjusted returns relative to passive diversification strategies. Daily data from November 2019 to April 2026 for four AI sub-sector ETFs (cybersecurity, cloud computing, blockchain, and robotics and AI) are used to forecast price direction 1 to 20 days ahead. The forecasts use tree-based ensembles and benchmark classifiers with a feature set combining technical indicators and business-cycle variables. Forecast accuracy varies by asset, model and horizon, and no single model is best for all ETFs. Tree-based ensembles reach accuracies of 60–74% at medium horizons and in most cases outperform logit and Naïve Bayes benchmarks. Tree bagging ranks high in accuracy over the medium horizon. Shapley values show that five-year breakeven inflation, the 200-day moving average, on-balance volume and Williams accumulation/distribution are the most important features, with higher inflation expectations and overbought conditions lowering the probability of price gains. A switching portfolio that holds ETFs only when 5-day forecasts are positive outperforms an equal-weighted portfolio after transaction costs, with a higher Sharpe ratio (5.65 versus 0.02) and a much smaller maximum drawdown (3.0% versus 26.2%). Stress tests show that a 20% market crash reduces the switching portfolio by about 5%, compared with 22% for the equal-weighted portfolio. These results are robust to different forecast horizons, trading costs, and rebalancing periods. Forecast-driven allocation can therefore provide downside protection that static diversification across AI assets does not.