I have been wondering how efficient a AI could be for trading the books at high frequency.  The overall market movement is very sporadic and prone to news, that will confuse most backtested algorithms.  But if you narrow down and focus on something like the orderbook as input, could it work?