Definition
Look-ahead bias occurs when a backtest or analysis uses information that would not have been available at the decision time. It can enter through revised fundamentals, publication timestamps, future prices, index membership or misaligned joins. The fix is a point-in-time information set and an explicit execution lag.
How it is calculated
A decision a_t must be a function of information available at t and must not include future outcome Y_{t+1}.
| Symbol | Meaning | Unit |
|---|---|---|
| a | decision or signal | decision-specific |
| t | decision time | period |
| g | decision rule | mapping |
| I | information set | set |
| Y | future outcome | outcome units |
Interpretation
A backtest must make each decision with the information that existed then, not the cleaned-up history available now. One leaked future field can make a rule look clairvoyant. The bias is about timing, even when every individual data value is accurate.
Assumptions and conventions
-
Information availability and timestamps are recorded point in time.
-
The simulation applies the intended order-to-execution delay.
Units follow the decision and outcome; t is a time index.
Limitations and common misreadings
-
Using a future value indirectly through a join is still look-ahead.
-
Point-in-time data does not remove other biases such as selection or costs.
References
- Marcos Mailoc López de Prado. (2018). Advances in financial machine learning. Chapter 11, The Dangers of Backtesting. isbn:9781119482086. Source

