Neural Networks Revisited

Great! We are on the same page. My testing confirms that is does NOT matter.

Here is the next "thought experiment". I am trying to predict the future price change FPC.
If I supply the FPC as an input, then the NN uses this information and makes pretty-much perfect predictions. No surprise there.

BUT now I want to see whether the NN can find a hidden relationship. So I have created 2 new inputs (InputA and InputB) like this:

for (int i = 0; i <=NumRowsOfData; i++)
{
if (IsEven(i))
{
InputA = Math.Abs(FPC);
InputB = Math.Sign(FPC);
}
else
{
InputA = -Math.Abs(FPC);
InputB = -Math.Sign(FPC);
}

}


Note that InputA * InputB = FPC.

If I give my NN ONLY InputA and InputB, do you think it will be able to discover the relationship and make perfect predictions again?

 

Lol...interesting. About how many rows. And how far into the future are we talking? One day? Twenty days?

 

5 days I am using at the moment. And currently I have 3000 rows.

 

Without knowing how many layers and hidden neurons, etc.; my guess is that the NN will split the baby and tend towards zero.

 

Well, I can choose to use whatever structure in terms of number of layers and hidden neurons. The question is do you think there is SOME structure that will allow the NN to pick-up this relationship? And, what do you suspect it might look like?

 

Not that I am aware of. My testing years ago showed that the NN would normally and simply output an average value rather than figure out some complex "wiring" of the neurons, when faced with strongly "zig-zagging" data.

I've never looked much into the whole DEEP learning and CONVOLUTED stuff. But my guess is that they too...would split the baby...maybe more so?.

Do tell the results if you run the experiment.

 

Oh, I should mention that I would probably split the problem between two or more SEPARATE NN's some kind of way. One would handle the "zig," the other would handle the "zag." I did find this to be the only way I could overcome this problem.

One NN will not be able to handle it, imo.

This is where creative pre-processing can make all the difference.

The more NN's, the better. Each will specialize in their assigned SIMPLIFIED duties.

 

Has anyone gotten lucky to see future proof results in stock market with neuron programming?

Overall it seems very impressive things have been done with images
http://karpathy.github.io/2015/05/21/rnn-effectiveness/
But does each task require specialized algos to input data to neural nets?
What sort of hardware is needed to test with neural network concepts?

 

NNs were great excitement 20 years ago BUT..........they never amounted to much. Can't fore-guess the market.

 

Just to be clear.

Because you haven't been able to amount much from NNs; and no has chosen to share with you that they've been extremely profitable with NNs; you've concluded, beyond any doubt, that NNs have never amounted to much?

And;

Since you can't "fore-guess" the "market;" you've also concluded no one else has nor can?

Thanks.