You are trading patterns, whether you acknowledge it or not. So, it's a strategy. Questions is, are you good enough programmer to "write it down". There are traders out there that analyize 100k+ strategies at a time, and make no mistake, your "discretional" trading "strategy" is perfectly found in one of those.
Let’s say my strategy gives me a short sell signal based on the pattern, but I decide not to take it. As reasoning, I give something like below: Time of day is past 11:30, we are in Santa Clause rally mode, above 50 and 200DMA, bad news are being discounted and other stocks in the industry are trading near all time highs. VIX is at the lows. Shorts haven’t been able to gain momentum all week. Decided to skip this trade as low probability of success. Can AI learn anything if I have thousands of such ramblings or do I need to quantify each point- like what exactly is “low” VIX, how do you define “Santa rally?” Etc?
What AI can do, is analyze 1000s of your executed trades and match them with existing patterns (and/or news events) and then come up with the closest possible strategy to mimic your behaviour. I'm kinda doing this but waaaay simpler than i described. There are several roads to reach the same destination. Your thinking is too complicated. The elegance (in everything, i guess) is hidden in simplicity.
there are tons of examples, discussions, competitions, ... of this on kaggle.com . Thing is that profitable discretionary traders (if they exist) cannot explain what they do so feature extraction could be useful. Tudor tried to do this with his discretionary trading but it got nowhere. That was many years ago
%% He might as well try it since he likes complicated stuff. EVEN if the complexity doesnt help\ the writing could help him a bit....... {LOL ''if they exist''}; they cant predict+ a system that makes stuff up sometimes , could ''get nowhere '' as you noted . But I just did some GOOG on Saint Nick rally+[used but did not type discretion] ; discretion was better than the machine definition of last 5 DEC + 1st 2 days of JAN=Saint Nick rally . ''I did get nowhere'' on the more complex made up definition [machine mentioned LOL] of any DEC rally being Santa Clause Rally LOL.
Yeah, AI could learn something as simple as that under a supervised learning model. You simply add another label to the model and give it a classification of 0 or 1. Feed it to the model. And you're done. Well, that's how I think I remember it from a class I attended over two years ago.
It depends.... If you can code a Heuristic algorithm, then yes. The key here is there are definable rules which the ML can learn. And thereby removing the learning process from an expert human to an Algo. If there is some pattern, then yes. So I would say, the first method, which is the route I have been working on for 4 years now, is possible but a lot of work. The second method, would most likely churn out many high risk, low reliability "system". So then the issue becomes, as with all ML how do you select the right system or parts. ML will learn something, the question is will it learn anything useful. And how will you verify it. It is really two problems wrapped into one with onion layers in each and co-dependent or orthogonally dependent. Hence why I am doing the first method. Lastly, just because you have some black or grey box churning out orders, does not mean you have a system. There is a lot of infrastructure that needs to be handled. It is a lot like banking transaction processing where you have things like "roll backs", "Try-Fails" etc. If you are interested in this stuff I would suggest you work on creating the harness for the orders that *would be* generated first. Get that to work. And work on the order generation at the same time. You need both. BTW: I have a background in this stuff since 1990.