Look-Ahead Bias

Look-ahead bias is what happens when a test consumes information that did not exist at the moment being simulated. The obvious version is a timestamp error. The dangerous version is quieter: a vendor silently overwrites a figure with its later correction, a fundamentals table is keyed to a fiscal period rather than to the date the filing appeared, or a universe is built from names that exist today. In each case the research is answering a question about a world that already knew the answer.

It is expensive because it inflates results in exactly the way that survives review - the equity curve looks plausible, the Sharpe looks defensible, and the error only appears when real money trades against a world that has not happened yet. The structural fix is not vigilance but data: a point-in-time archive that records what was known when and never rewrites it, a universe that keeps its delisted names to avoid survivorship bias, and features computed against event time rather than against a period label. A backtest is only as honest as the history underneath it.