prediction always 0

[deanshub]

Hi,
Iv'e altered your code to support google API instead of yahoo which stopped working
but the prediction seems to always return 0, even the acc prints always 0,
can you please explain why?

[zhivko]

Exactly, I did same (changed google for yahoo to retrieve data), and this is what I get:
It would be nice if you could comment how to fix this.

You can check my source code under:
https://github.com/zhivko/lstm_stock_prediction/blob/master/.ipynb_checkpoints/predict.py

[etai83]

Hi guys!
try to do the following:

1. change the stock value in the df object to contain 0.3 values (divide it by 1000 instead of 100).
2. try to minimize the window parameter value from 22 to 5.

This will help the model to train more efficiently.

[etai83]

tell me how it worked

[zhivko]

Tried changes you suggested - but result it same (same graph as above), however I can see from output minimizing is happening. Check output from https://pastebin.com/AXXiBX41
Maybe problem is just in charting?
Check updated code at:
https://github.com/zhivko/lstm_stock_prediction/blob/master/.ipynb_checkpoints/predict.py

[etai83]

Your submitted code does not contain my first suggestion. Try to divide the Close, Volume, High by 1000.

[zhivko]

Oh - that's better.

I pushed the code to my repo.

[zyf3883310]

this paper I saw, but this solution is still not possible

[zyf3883310]

so, can you provide a good solution

[zyf3883310]

or is the stock model not good?

[etai83]

Try to run the code's latest version. I've uploaded one yesterday.

[zyf3883310]

hello why this input . such as stock high,open,close must divide 1000?

[and-rewsmith]

Why are you dividing close/high/output by 1000? Mine flatlines if I don't divide but I don't understand why. I've tried scaling and using leaky-relu instead of relu but that did not fix it. Any reasoning behind that?

[etai83]

In this example dividing by 1000, we are helping the algorithm to learn more efficiently and get better results much faster.

[and-rewsmith]

Why does that help the learning efficiency?