Avoiding Look-Ahead Bias, Over-Optimization, and Curve Fitting
Summary
The article describes three ways strategy research can produce misleading backtests: look-ahead bias, excessive parameter optimization, and curve fitting. Its examples show how using a bar’s eventual close to trigger an earlier trade, or assuming a breakout fill at a price reached before the breakout was known, can give a strategy information or execution prices unavailable in live trading. It also explains signal flicker, where an intraday condition disappears by the close after a trade has already occurred.
To reduce overfitting, it recommends limiting the number of parameters and rules, checking default and varied parameter settings, and evaluating strategies across multiple instruments. It warns that rules tailored to historical periods may fail when market cycles change. These are practical research cautions supported by illustrative examples, not empirical tests or a formal validation framework. The discussion does not specify statistical procedures for choosing parameter ranges, separating training and test data, or modeling transaction costs beyond its slippage example.
Key ideas
- Look-ahead bias occurs when a backtest uses information that was unavailable when the simulated trade would have been placed.
- A breakout test should account for the price movement and slippage needed to enter after the breakout is observable.
- Signal flicker can make an intraday signal vanish by the close even though a live trade has already executed.
- Using many optimized parameters can make a losing strategy appear profitable in historical tests.
- Simple rules, varied parameter checks, and testing across instruments can help expose fragile strategies.
Tags
This summary was written by Stratmill's research agent from the original; it is not a copy of the source.