Using Machine Learning for discretionary trading

[qlai]

Let's assume you are a consistently profitable discretionary trader (or have full access to his/her trades and, most importantly, train of thought).

Would it be possible to use machine learning tools to teach a bot to trade the same way? Let's say you can code basic trading and money management rules, but need to add context related filters which cannot be easily put into code (the discretionary part).

If we are talking about day trading (3-6 trades a day) would it take just a few years to get close enough?

 

[tsfx]

Let's assume you are a consistently profitable discretionary trader (or have full access to his/her trades and, most importantly, train of thought).

Would it be possible to use machine learning tools to teach a bot to trade the same way? Let's say you can code basic trading and money management rules, but need to add context related filters which cannot be easily put into code (the discretionary part).

If we are talking about day trading (3-6 trades a day) would it take just a few years to get close enough?

It depends how "random" is your discretionary strategy. But i'd say it takes more than 1000 trades as an input to match with a quantifiable strategy. Unless you're discretion is event driven. The more tech driven the easier.

There's really simpler ways to create a strategy

 

[qlai]

There's really simpler ways to create a strategy

Thanks, but it's not about creating a strategy, rather kind of a skill/brain transfer.

 

[2rosy]

I think you could create the probability of a certain action taking place given certain features at the time. There are off the shelf ML algos that do this. If you have the data it wouldn't take a few years. Maybe a few hours or days to see if it's worth your time

 

[ph1l]

This might be possible with supervised learning having

• inputs as various indicators and relationships (e.g., price at current bar is greater than price N bars ago)
• outputs as parameters for a trade (e.g., entry relative to most recent price)
• fitness function that compares the outputs for a particular trade with an actual trade at or near the simulated time of the outputs (e.g., measure how close an output is to the actual trade with closer being fitter)

It doesn't sound too easy to do though.

 

[notagain]

Teach your bot to stand on the beach of Taiwan or shadow Nancy P.

 

[murray t turtle]

Let's assume you are a consistently profitable discretionary trader (or have full access to his/her trades and, most importantly, train of thought).

Would it be possible to use machine learning tools to teach a bot to trade the same way? Let's say you can code basic trading and money management rules, but need to add context related filters which cannot be easily put into code (the discretionary part).

If we are talking about day trading (3-6 trades a day) would it take just a few years to get close enough?

%%
Sounds like you could,mostly, dont know about your time estimate.
But since cant predict+ finance is mostly behavior not about ''most importantly train of thought'' go for it anyway+ oversee it anWay [discretion].........

 

[engineering]

The question doesn't even make sense.
If you have a straightforward strategy you can write down, you don't need AI you just need to white a normal computer program to follow it.

The point of AI is to try and do things it isn't explicitly programmed to do.
If you already had a consistently profitable strategy, why would you even want to computer to be able to deviate from it?

 

[qlai]

If you already had a consistently profitable strategy, why would you even want to computer to be able to deviate from it?

I’m not trying to deviate from it, I would like to automate it and possibly transfer it to future generations. I’m looking for a way to transfer skill which has lots of discretion. I’m not looking to improve existing or develop a new strategy. Can AI be used for such purposes?

 

[BlackPhoenix]

You would first need to know what inputs the trader uses for the trading decision and try to minimize that input. E.g. if the trader does naked trading purely based on price action without outside influence or if there's fundamental analysis or news that influence the trading. Then you could feed that minimal input to a deep learning network to produce the same results as the trades made by the trader. I would be amazed if someone hasn't tried this already though