Multi-Horizon Machine Learning for Stock Selection

Summary

The report describes stock-return prediction with machine-learning models built for short, medium, and long forecast horizons. It groups selection factors according to their information coefficients at different horizons, reflecting the idea that factor information decays over time and that short-term sentiment signals may suit shorter forecasts while persistent price distortions may suit longer ones.

The reported results suggest that the best model depends on the portfolio’s rebalancing frequency: short-horizon models suit more frequent turnover, while longer-horizon models perform better with slower turnover. Under an assumed trading cost of 0.3%, weekly rebalancing is reported as more suitable than monthly or quarterly schedules. Because signals from the three horizons are described as weakly correlated, the report also combines them; the weekly ensemble is reported to outperform individual models and have lower turnover than the short- and medium-horizon models. These are summarized findings from the report, and the excerpt does not provide methodology details needed to independently assess robustness or generalization.

Key ideas

• The report builds separate machine-learning stock-selection models for short, medium, and long forecast horizons.
• It groups factors using their information coefficients across forecast windows and notes that predictive information decays over time.
• The reported evidence links shorter-horizon models with frequent rebalancing and longer-horizon models with slower rebalancing.
• The excerpt reports weekly rebalancing as preferable under its stated transaction-cost assumption.
• Combining weakly correlated signals across horizons is reported to improve performance and reduce turnover relative to some individual models.

Tags

This summary was written by Stratmill's research agent from the original; it is not a copy of the source.