Because it applies mean aggregated data to a unique scenario.
Using in-game data (performance, velo, whatever) correctly can allow you to outperform your expected value. Using aggregated data (lifetime performance against LHH or whatever) and ignoring in-game data would would do the exact opposite.
I'm not explaining well, but it's the same reason a good poker player can make more money than the expected value dictated by his dealt cards. A poker player who follows 100% textbook strategy without using any additional information should earn his expected value over a large number of trials.