[rl3]

With that username, I suppose humanity should be thankful that OP isn't applying for a job in AI. :)

On that note though, I'm curious how useful modern deep learning techniques are for trading. Traditional machine learning for the past couple decades was never terribly useful for entry or exit signals, merely as pattern recognition for ancillary purposes. I'm wondering if that's changed now.

If I was to guess, ML is probably easier applied to longer time frames. On the other hand, it's certainly possible to pipe the output of a GPU supercomputing cluster to an FPGA/ASIC HFT fleet. A slight delay in updating the trading model isn't the end of the world if execution speed remains good.

[AppeasingRoko]

My not particularly informed view is it wouldn't be great. Overfitting would be the biggest problem, as well as the fact that people design trading software specifically to exploit other trading bots these days.

On the plus side, less human input in trading software may greatly reduce human biases which make us suck at trading.

[rl3]

>... as well as the fact that people design trading software specifically to exploit other trading bots these days.

I figured modern ML would be the first thing firms use for that purpose. Might be interesting though with bots trying to learn the behavior of their counterparts and vice versa. The ensuing feedback loops are probably quite interesting.