search

nateemma / strategies

Custom trading strategies using the freqtrade framework
329 stars 88 forks source link

The issues ahead... and the logical solution!? #17

Open just-nilux opened 1 year ago

just-nilux commented 1 year ago

So your work is really incredible, I've build a lot of custom strategies out of this repo. However, when the strategies are using some more complex models the limitation of Freqtrade are obvious.

I think you running into a lot of unnecessary problems that would be simple to solve by just using FreqAI. I've got into it a few months ago and the recent progress thats being made is exciting.

The most serious issues that I see can't be resolved without core changes to Freqtrade, like:

I wonder why not migrate all these great classifiers you have already for FreqAI? I started with Anomaly (Autoencoder) already and want to make it work with PyTorch MLP of FreqAI. Almost done...

I'll see where I can help and make a PR. But it would be good to know where you want to head with this, so the right efforts can be made. Feel free to get back to me on Discord (nlxtrading#4404) or Telegram (@nlxtrading) if you want to chat. Thanks for all your hard work and good luck man! ...

tomjrtsmith commented 1 year ago

Sounds like a good idea to me. I checked out your repo nilux nice work with the Leledc bars.

I'd be in favour of maybe asking Mathais if a stable version of Tensorflow can be bundled in with the Freqai install along with some of the other packages (pywt doesn't seem to ever change). That would help the TF bugginess.

just-nilux commented 1 year ago

@tomjrtsmith this PR deals with the tensorflow topic and I've tested it, works well. https://github.com/freqtrade/freqtrade/pull/7807/files

Another huge roadblock for the strategies as they are right now is feature engineering. With freqtrade you can't really do any feature engineering and it is such an important part of building ML/AI strategies. Especially when working with future data you want to control exactly what is available to the model to reduce lookahead....

tomjrtsmith commented 1 year ago

Surely there's a way to adapt all the classifiers and anomaly detectors to use Freqai without rewriting everything for PyTorch?

just-nilux commented 1 year ago

Surely there's a way to adapt all the classifiers and anomaly detectors to use Freqai without rewriting everything for PyTorch?

Well, I can see why it's more fun to build such a framework from scratch... especially when doing research. And no, I don't think everything is possible in FreqAi yet cause of the dependencies that need to be resolved. So yes that's an issue if you want to develop and experiment with different algorithms and estimators and such...