Trading data model architecture

Suppose we have loads of historical trading data with all the important features associated to it. How do we start to design a model to predict better trading decisions (excluding market sentiments, just by previous 50 years data)?

This is an extremely difficult problem.

Decades of attempts have shown that past performance isn’t predictive of future performance, because the training data doesn’t provide any information about future unforeseen events.


@TMosh - Thanks, that makes sense! But is it possible to identify different patterns(generally indicated by graphs by traders) from the historical data?
Or maybe in some way helps in decision making process?

There’s lots you can do with making plots of historical behavior.
They’re just not very useful for predictions.

There is another thread that discusses this topic, I’ll post a link if I can find it.

Here it is, in the AI Projects area:

Thanks @TMosh !

@Blaze , let me know if you want to discuss this. @TMosh is correct that this is very difficult to do.

Just as a counterpoint to TMosh’s point, I know it’s possible because I actually had a bot that did very well at predicting the movements of crypto currency for awhile. We had a bot that earned around $300K/month on a portfolio of $1.5m by making one leveraged bet each day on the direction of BTC and winning each day’s bet. It did this for about 3 months before getting liquidated and losing about 70% of the portfolio on one bad bet because the engineer hadn’t built sufficient protections into it.

My problem is that this was from code that another engineer developed for me. And we parted ways on bad terms. I’ve been trying to reproduce it on my own for the last year and have experimented with different technical patterns.

That’s relatively easy in a rising market.

This is not an uncommon outcome.

Actually, it was in a declining market because it happened during the summer of 2022. And the majority of the bets were short bets but it also sometimes made long bets as well.

But ultimately, it’s unfortunate that it got liquidated on one bad trade. :cry: :frowning:

@TMosh , here’s one question. How do you quote only part of a comment in your replies? Thanks!

Click and drag to highlight, then release and you get a popup that has a “Quote” tool. It automatically pastes the copied text into whatever message box you have open.

Disregard of accuracy of trade model, can someone please give directions into what ML architecture to use to start with?

Bitcoin price and price volatility historical data

Given that kind of price volatility I’m not sure what ‘rising’ and ‘falling’ market even mean. It’s not happening on a ‘business cycle’ kind of time scale, that’s for sure. For crypto price modeling can’t one throw out most of the factors that complicate traditional asset pricing and portfolio optimization - the kinds of things a pension has to do to offset its long term obligations, for example. Seems like it reduces to guessing right on price momentum on a time scale of minutes to hours, which is a different kind of problem.

If you allready have historical data , hourly not daily , I can try some fetures using my supper energies then go with an AI model on it