On the predictability of ETF returns with technical predictors

We develop a machine learning framework that uses technical indicators derived from international stock data to forecast the performance of international equity ETFs. To address the limited history of ETFs, we train a random forest model on global stock data and apply it to rank ETFs according to the probability of outperformance. Portfolios formed on this ranking are associated with economically meaningful gross-of-cost return spreads, averaging 0.76% per month (t-stat = 2.76) and 0.62% after market risk adjustment. Predictability is strongest at short horizons and decays with longer holding periods. Volatility and momentum indicators contribute the most to model performance. In line with limits to arbitrage, predictive strength is more pronounced in less efficient markets and among lower-liquidity ETFs. Results are robust across portfolio construction methods and alternative models, and the signal performs well out-of-sample from 2011 to 2022, with extended evidence up to 2024.