I have recently compiled what I consider to be the 10 most important rules of using non-linear models, especially neural nets. I will be going over these in detail at the NeuroShell Trader Seminar, but except for the obvious financial ones, they apply to science and business modeling, too.
Sincerely, Steve Ward
Ten Commandments of Non-linear Modeling and Neural Nets
Copyright 1999 by Ward Systems Group, Inc. These commandments may not be reproduced or distributed without permission of Ward Systems Group, Inc.
1. Thou shalt seek inputs that affect the output.
2. Thou shalt not use too many inputs.
3. Thou shalt not use too few training patterns.
4. Thou shalt choose thy training patterns so that they are not clustered and are a balanced representation of the problem.
5. Thou shalt not judge thy model by its ability to reproduce training patterns.
6. Thou shalt not apply statistical measures and precision to financial modeling.
7. Thou shalt not use too many categories when classifying.
8. Thou shalt not code inputs with monotonic values unless they represent monotonic concepts.
9. Thou shalt not waste time coveting thy neighbor’s financial models.
10. Beware of extrapolation of non-linear models.