Look-Ahead Bias
Using information in a study of the past that was not actually available at the time, which makes results look better than reality allowed.
The bias enters through small, unglamorous channels: a supply figure that was later revised is treated as though the corrected value was known at the time, exchange labels created after an investigation are applied to flows from years earlier, and a value published hours after the moment being modeled is used as if it were available at that moment. A related version appears in the choice of universe, when a study covers the assets that exist today rather than those that existed on the historical date. Digital-asset data is especially exposed because so much of it is backfilled and so many labels are retroactive. The corrections are mechanical: use point-in-time data vintages, lag every input by its real publication delay, and record when each piece of information first became available.
In practice
Computing historical exchange flows with today's address labels assumes an analyst in the past knew wallet attributions that were only established later.
The common misunderstanding
Look-ahead bias is usually accidental, creeping in through revised data and retroactive labels rather than through deliberate manipulation.