As Bowgett mentioned and (with my little experience) I also agree... I think neural nets are mostly useful for on-going parameter optimization vs using them to make trading decisions. What I do and suggest to others.... I would focus on building profitable strategies first then see where a NN could help with optimization if its needed. With as much computing power is available today, from Amazon EC2 and others, you should be able to complete optimization for hundreds of instruments by the end of the weekend for the coming trading week.
optimisation looks like the Holy Grail but they never go on performing as well. Mathematicians can calculate the orbits of comets and whole galaxies but they can't forecast the markets yet. Indicators use past historical data and for that reason lag badly. Barely better than a coin toss. Even clever people get seduced by the huge amounts of money on offer and can waste decades and fortunes trying to find the correct answers.
Here is where the entire logic of the experiment falls apart. Any result that passes your test will be nothing more than a curvefit..simple survivorship bias that is not logically rooted in any market principles What i've noticed about people who are interested in NN, and GP, and machine learning is that "playing around" with the TECHNOLOGY is more important to them than the actual act of TRADING and making money. You have a free machine that learns..it's called your brain
My experience is 1. NN's tend to immediately over-fit the data. 2. It's hard to come up with a performance function to train the NNs. 2. You may end up not trading an NN alone since you will not trust it. But they are great to play with!