I think statistical methods are the best primary option. There are a lot of other tools you can use but this is the most impenetrable from the outside.

The chances that the cheater is able to anticipate the statistical state of everything logged server-side is negligible. There is no way to "sandbag" performance on purpose if you don't know how your performance is being measured.

There is also the problem (solution) of sample size. The players' performance in one or ten games is ideally not relevant to the heuristic. There is a threshold that is crossed after hundreds of rounds of dishonest play. Toggling cheats within a match or tournament series would be irrelevant.