Sorry, 20 period change then. I'm guessing you are not "sorting" the prices, but just making it reverse. How? Do you make the numbers negative, or do you divide by the factor between periods? If there's differences, it could be explained by how you do it. If you make the change in percent, couldn't you just flip the sign? It's a bit different than most people do it, but I make it a point to always try to be numerically unbiased as towards direction. These are types of things an NN won't help with directly. Update: If you're reversing the bar order, why would you expect same performance?
I am just reversing the bar order - no other changes. And Stock A is in a strong rally through the period, and therefore Stock AR is in a strong decline. I would expect the same performance because the only thing different is that the sign of the changes I feed to the neural net have flipped.
If the NN was trained on mostly upward moving prices, it probably won't fully understand downward movement even if it is just "flipped". I've heard of some people using one NN for price increase and swapping to another for price decrease. I can't remember if it really outperformed a single NN, but it's an interesting concept.
I train on the first part of each dataset independently, but with the same inputs (except reversed data of course)
So you basically have a one input, one output NN. Input is 20 day price change, Ouput is the next 20 day price change...for the forward NN. For example, it could look like this: 10,11 11,13 13,9 9,8 8,6 etc. Is that correct?