Robust Short-Horizon Equity Models

Article BigQuant

Summary

This practitioner note compares very high-return, moderate-return, and lower-return backtests for short-horizon equity models, focusing on whether their apparent performance persists out of sample. The author reports that aggressive timing filters and models trained on short histories often lose effectiveness quickly, while longer training histories are described as more robust across changing market styles. The note also argues that simpler models with fewer factors may generalize better than models with many inputs.

Its evidence is an informal account of the author's tests and live or simulated experience, not a controlled study; detailed datasets, evaluation procedures, and comparable performance statistics are not supplied. The article cautions that fixed-schedule rolling retraining can make a model worse, and that high backtest returns alone are a poor selection criterion. Its recommendations are observations for short-cycle models rather than established rules, and some stated preferences—such as favoring high returns—sit uneasily beside its warnings about overfitting.

Key ideas

The author reports that heavily filtered timing models can have short-lived out-of-sample performance.
Short training histories are associated with sensitivity to market-style changes.
Longer training histories are presented as potentially improving robustness, based on informal observations.
The note favors fewer factors and stable alpha, while warning that larger feature sets may overfit.
Scheduled retraining can introduce instability when new data weakens a model.

Choosing Robust Short-Horizon Equity Models Overfit to Backtests

Summary

Key ideas

Tags