(10-11-2019, 08:24 AM)Dieguz Wrote:(10-11-2019, 08:23 AM)Dieguz Wrote: Hello,
i have found a cool CNN implementation in tensorflow.js applied to stock markets.
https://github.com/Gago993/StockPredictionCNN
i think it is a good starting point to learn how to code in Tensorflow and adapt to crypto market.
PGTART: if you feel that it is a good implementation of CNN applicable to the crypto world, i can do my experimentations to adapt it to gekko or rewrite it for that.
Diego
Well I looked quickly in the code.
At last this is a real neural network
It got multiple layers (model add)
It uses 7 inputs (a bit small number)
Not sure what is put in as data input though.
But this code might be rewritten.
For example feed it with various tulip counters.
It seams a good starting point for coding.
Just a few notes:
Neural nets need to be trained the training method has great influence as well.
Training takes time and the more nodes in all the layers the more time.
It takes some NN experience to optimize a NN don't worry about it i or others can do statical analysis of the model
You might not require a time series NN model
Those are usually not that good in predicting. The good ones score around 80% while normal nets easily go 96 and higher.
One needs to be able to save and load the weight model of the neural net.
(I can write it if it's not there)
You will need a series predictor if you feed it with Manny candle bars. You won't need a series predictor if you feed it with other indicators (be sure to scale their values in-between -1 and 1
Use relu and use a bottle design ea many input nodes per layer use gradually smaller dense nodes in end layers. Such networks are usually good there is a lot theory about it of why whiches down to slowly converging of information density.
If you got a basic tensor flow neural net working we can then later create a good layer model for it
The code might serve as a hello world tensorflow Gekko. Don't worry if it won't go super. I codes lots of TF NN we can tweak such a network later a lot. Also saving and loading I can help or write.
We could write multiple NN based on a first I've written so many nets it be fun
Send me a pm if you got a basic working version.
[....Resistance is futile...]