Sorry, if it sounds boring. But I am trying to make my way through AI usage in trading on a meta-level (or metho-...).
There are several questions that might be worth discussing.

1. This one has already been touched. Why would AI be a (better) way to go in future? The short answer was: because if we agree that mechanical systems can earn a return then probably self-adaptive ones are even better because the markets change.
While I don't argue that AI systems can be better at adapting, I have doubts about "changes". It seems to me that even with huge computational power AI systems will only adapt within certain opportunity space with fixed dimensions, whereas markets can change in more drastic ways. And we should still have a human tuning the system manually in some ways, in addition to the automatic ways embedded into the system.
For example, we have a system that reads N last ticks and encodes them into some vector that is fed to the system. System can adapt and change withing these limits - N-dimensinal space only. Now, what humans can do - they can change N, change the way system learns, change encoding scheme and change the way we look at the market completely. Humans may decide to take 2N ticks in order to extract N random ticks and deal with them from now on. How's that?

3. How to tell if the system is going to work on out-of-sample in the first place.
My understanding of this topic so far - we split the problem into two: A) check if Ai system adapts to some benchmarks successfully (I don't know - some trigonometric curves, random walks, fractals), B) statistically test price time series for presence of the processes that the system is supposed to adapt to.
Since the item 1, we cannot hope that AI system is going to adapt to anything and everything, we constantly have to model what we find in time series in order to produce more and more tests for stage (A).
Or does anybody have a different view?

Is this topic of interest to anybody at all?

Hi DerikB,

I agree with you there. One does not have to be a genius to construct an efficient regression model. What takes tremendous amount of time and effort is to come up with a model (structure) that can be "trained" and perform well on out of sample data.

The old argument stands: if training adaptive models was easy, everyone would be doing it.