While I believe that machine learning will change the world in finite reasoning domains (judges, lawyers, driving, doctors, teachers, etc...all whom will lose their jobs soon to these machines) from my point of view there is one aspect of it that leaves it wanting in infinite adaptable domains.
Imagine that a machine was trying to prove the Riemann Hypotheses, and it came up with a proof that looked like this:
(0111,333,22,111199977752314159, 6947,...900 trillion numbers later, QED)
Every machine in the world, using whatever theory was used to construct such a sequence, verified that the Riemann Hypotheses was indeed proved. What good would it do other than the proof itself?
The problem is that computers, at least as we understand them today, compile down. Human beings compile up. What I mean is that computers just build an incomprehensible assembly language of correlations with finite nodes to make inferences. Whereas human beings compile up to general emerging principles . It is hard to state, but what I think I am getting at is that ML and computers in general are trying to prove theorems, while human beings are going up and down the reasoning stack simultaneously, and most importantly, intuiting General Theories that don't exist at all in any of its finite systems of current understanding to emergent general principles of reasoning. Said in another way, human beings, particularly artists of which I include mathematicians, are building higher level Godel models/languages to make statements in lower level languages vastly easier to understand.
Imagine the following situation in applying a similar ML technique to trading. Say you build a trading system using the same technology as used by Googles DeepMind, or something even more advanced like the system above that goes on to prove the RH. Imagine that you spend a huge effort to make a system that seems to work, from the point of view of profitability, but you don't understand a single reason as to how it is making decisions. Still, you are brave and you turn it on live. The system does great for a couple of months, then it is breaking even, then it starts to lose. What do you do?
So, the problem with ML imo particularly as it applies to trading, is that since the human being cannot understand the output of the machine, and since the ultimate node in a system is the human being deciding whether the system was viable or not after some adverse run, it leaves us in an anxious state both while running it and while deciding to take it off line.
That is not a trading foundation to build on.
