I know this thread is pretty old already and this is my first post here BUT when I searched on Forex NN I came here I found the reading pretty interesting but I can clearly see most of you approach this 'approximation problem' from the forex side and might (or might not) see Neural Networks as a 'holy grail' to forecasting. Actually it can be but that aside.
First of all, I am (only) 25 and have about 1 year experience in forex trading(like from 3 years ago when I had time to much). But now I do a Masters study in Intelligent Systems and with that AI like ANN are of daily order.
Did anyone ever wondered about the actual structure of the NN? I mean:
- Does it contain hidden layers (like a must for approximation!!!) if yes how much? 2 hidden layers(its is considered 1 could approximate almost all functions) could be possible for approximation since financial problems are really complex.
- If so what learning algorithms does this NN use? Standard back propagation? Mathatan learning rule? Resilent training? or maybe even: levenberg quadrant?(this would be for candle sticks rather than EMA).
What activation functions are used? naturally if you do not normalize the inputs it would be linear and output aswell but what was used for the hidden layer?
How are the inputs selected? Just by trial and error? or based on personal experience?(not bad though) or are ALL possible inputs tried and these came out as the best? (in matters of checking the NN weights and see if the NN ignores them as they provide no good information)
If there is some answer on those questions I would LOVE to know them!!!