Curious about the code

[davecramer]

@elfayoumi

What is the point of doing the difference between the two observations?

Is there somewhere I can find your slides for the talk ?

[elfayoumi]

Hi davecramer
The example is trying to use the technique to predict the future gain, so from today's past observation, will ASX value increase, decrease, or stay the same?
Financial data are not stationary and as in this example, we could not get more than %39 accuracies, although it is higher than %33 of random guess, it is still close.
So in this example, I am not trying to predict the future value but the change, which is difficult, and the most important.

Sorry, there were no slides, I just used Jupiter notebook to have all the required information.

thanks

[davecramer]

thanks. I am python challenged so managed to muddle through it. I'm curious that I don't see any large NN libraries including GAM, is there a better way to predict signals ?

[elfayoumi]

I think the main thing is it increases the size of your data, you convert one dimension array into a two-dimension array, and something like RNN is sufficient, but the idea, is Convolution network has the best performance now, so it is an idea you may like to try, it may work in your case.

[davecramer]

where does nn come from? I don't see the import ?

[elfayoumi]

Hi Davecramer
It comes from pytorch
https://pytorch.org/docs/stable/nn.html
fastai imports all relevant imports.

[davecramer]

interesting. When I try to run it using python3 it fails.

As you can tell I'm not particularly comfortable in python. Much more comfortable in a JVM language

[elfayoumi]

I need to update the code, this was fadtai version 1, we now version 2, pytorch is 1.8
you can add
from torch import nn