This is always the problem with training from historical data only: you’ll become very good at being just as good as the sample group.

Ideally you just use the sample as the basis, and then let an AI engine play against itself for training, and/or participate in real world games, such as they did with AlphaGo and/or AlphaStar.