It appears whenever the set you study was filtered by something related to performance. Testing a strategy on the ten most liquid crypto pairs of today means testing on the ones that survived and grew. Reading a broker's list of top traders tells you about the right tail of a distribution, not about the strategy.
survivorship-bias is the best-known special case, but there are others: instruments that only have clean data because their issuer stayed in business, brokers whose fill data you have because you stayed with them, and strategies you remember because they worked.
Example: measuring the performance of tradeable ETFs listed today across 2010 to 2024 excludes the roughly one in four launched in that period that closed, and those closures were concentrated in the worst performers.
Related: survivorship-bias, restatement-bias, data-snooping