Tom, I agree with your comments on using standard networks. The networks
that I use are far from standard. These nets have logical branching and adaptive
functional creation. There are 63 inputs that are computed from the data types
that I showed in your previous post. These inputs are moving averages and basic
technical and fundamental indicators. The networks take these 63 inputs and internally
generate functional forms that best model the market over about 5000 days of
data. This data is the training set. The most recent two hundred days of data
are not included in the analysis and therefore provide a blind test set. The
functionals that are created are very complex and I don't fully understand them.
Each network in the library of some 200 is different which increases the complexity
of the overall system.
If you need more
in-depth technical info, it might be best to email me.