We reliably use statistical process control to automatically calibrate incredibly precise, nanometric-scale machinery for purposes of semiconductor engineering. Surely, with the extreme amount of data available regarding every player's minute inputs in something like a client-server shooter, you could run similar statistical models to detect outliers in performance. With enough samples you can build an extraordinarily damning case.
The only downside is that statistical models will occasionally produce false positives. But, I've personally been "falsely" banned by purely deterministic methods (VAC) for reasons similar to others noted in this thread (i.e. leaving debugging/memory tools running for a separate project while playing a game). So, in practice I feel like statistical models might even provide a better experience around the intent to cheat (i.e. if you aren't effectively causing trouble, we dont care).