Look-ahead bias
What goes wrong when a past decision, or a rule tested on past prices, is allowed to use information that was not available when the decision was made?
In plain words
Look-ahead bias is letting data that did not yet exist shape a past or simulated decision, then treating the result as achievable. It matters to anyone running a backtest (a test of a trading rule on past data) or reviewing a track record, because it makes results look better than was possible.
Why it matters
It turns hindsight into false skill. A backtest, rating or review that quietly feeds a decision with later prices, highs or funding rates (payments between long and short holders of a perpetual futures contract, a contract with no expiry date) will make ordinary choices look brilliant. Rules built this way tend to fail once they are run in real time, when the later data is not there.
In a real week: ETH, week of 28 December 2020
ETH opened the week of 28 December 2020 at 685.1 and closed at 978.28, a 42.79% gain. Measuring a decision taken at Monday's open against that later close is ordinary evaluation: it records what happened. The bias enters when the decision itself uses later facts. A simulated rule that buys at the week's low of 681.04 relies on knowing it was the low, which could only be confirmed after the week ended. A rule that targets the week's high of 1011.07 relies on a price reached only at 23:00 UTC on Sunday 3 January.
A common mistake
Treating any comparison with later prices as look-ahead bias. Scoring a past choice by what happened afterwards is legitimate evaluation. The bias is letting later information decide when to act, at what price, or which cases to include. Human reviews, not only computer models, can make this error.