Let me tell ya something, guys...
1. ANN's don't work, they are no magic solution to to the 'almost' efficient markets. I have experimented with these a lot and I have almost abandoned them as a sole decisioning algorithm. They can still be used for subproblems though .
2. No matter what algorithm you use, your success will be driven by your predictor variables. The improvement of your results that can be attributed to the statistical/machine learning framework will be marginal, well, in most cases one will try a few different approaches and choose the best...
3. A successful system will require many steps- let's say we may use spectral analysis to detect cycles, wavelets to detect short term trends, other statistical measures as inputs to a NN...
4. In the end, in my opinion... we CANNOT predict shiat 90% of the time, we can only hope to be mostly correct in the remaining 10% when we dare to trade.

Cheers!