Automated stock picking using random forests

We derive a stock ranking by applying a technical features-based random forest model on an international dataset of liquid stocks. Rather than predicted return, our ranking is based on outperformance probability. By applying a decile split, we find that long–short portfolios achieve Sharpe ratios of up to 1.95 and a highly significant yearly six-factor alpha of up to 21.79%. Moreover, we show that outperformance probabilities serve as a superior measure of future returns in the context of portfolio optimization. Mean–variance portfolios using this measure are less volatile and more profitable than equally- or value-weighted portfolios. Our findings are robust to firm size, regional restrictions, and non-crisis periods and cannot be explained by limits to arbitrage. • Technical indicators predict outperformance in the cross-section. • Investors may identify outperforming stocks using machine learning models. • Outperformance probability is a superior measure of future stock return. • Identified return patterns remain valid for large stocks.