search

krystianity / coinpusher

:chart_with_upwards_trend: real-time cryptocurrency chart prediction based on neuronal-networks :bowtie:
Mozilla Public License 2.0
166 stars 49 forks source link

Prediction below 0? #3

Open imkane opened 7 years ago

imkane commented 7 years ago

Not sure if this is normal or not?

screenshot - 11_16_2017 5_00_41 pm

Maybe it just needs more time for training? Or maybe a dataset size change is needed?

krystianity commented 6 years ago

This looks like unlucky or not enough training data :) To get usefull graphs its a good idea to let coinpusher collect for a few days. Also to improve the results, I suggest working on an improved feature vector size, it will only detect patterns in the historical graph data, to get better predictions it definitely needs more data e.g. twitter feeds, crypto news, newspapers in general etc..

bitcoinvsalts commented 6 years ago

I have the same issue. @imkane did it get solved with more training data? @krystianity are you still working on this repo?

krystianity commented 6 years ago

@jsappme This repo is a snapshot of an early version of this project. Ment to inspire others and show them possibilties, there wont be any more commits from my side.

bitcoinvsalts commented 6 years ago

@krystianity Thank you. Indeed this repo inspired me a lot. I took the best parts of it and added it to my own custom trading platform. I trained a few LSTM networks using your code and I had some interesting results on the NANO/BTC pair on Binance. But for now, it's still too incoherent, I cannot let the AI run the show yet. I mostly use the AI for my dynamic trade exits (and risk mgmt).

What's your point of view and analysis of using neural network like lstm for price/time series predictions? Are you still using lstm in your current version of the project?