Screening Automatically Generated Trading Strategies for Robustness
Summary
The document argues that computing power can generate many candidate strategies from combinations of factors and rules, but that these candidates need research review before use. It outlines a screening process focused on avoiding look-ahead bias, detecting overfitting, judging when a strategy may have stopped working, and accounting for trading frictions. Suggested checks include rolling or forward analysis and paper trading, parameter-sensitivity analysis with human review, maximum-drawdown-based invalidation, and comparing predicted with realized information coefficients.
For implementation realism, the process calls for including commissions, bid-ask spread, market impact, and slippage. These are presented as evaluation practices, not as tested results: the available text gives no strategy examples, thresholds, numerical findings, or evidence that the proposed checks guarantee robustness. The core lesson is that automatic strategy generation is only a starting point and that validation must address both statistical reliability and actual trading conditions.
Key ideas
- Automated combinations of factors and rules can produce candidate strategies for researchers to evaluate.
- Forward analysis and paper trading are proposed to help detect look-ahead bias.
- Parameter sensitivity and researcher judgment are proposed as checks on overfitting.
- Maximum drawdown and predicted-versus-realized information coefficients are suggested for judging strategy validity.
- Commissions, spread, market impact, and slippage should be represented when assessing trading costs.
Tags
Cited by
This summary was written by Stratmill's research agent from the original; it is not a copy of the source.