Input/Output Data Convertion in BackProp.

[Scott E. Fahlman]

    	I'm using neural networks to predict financial time series, 
    specifically the gold-prices' time series. 
    
    I have mainly used LINEAR CONVERTION (original data value converted to 
    the ranges 0, 1 or -1,+ 1), but it does not seem to work properly.
    
    My results for the single-step mode are good, but i think it could be 
    better.
    
    On the other hand, for the multi-step, the results are very bad.
    
With so little information about your problem or the architecture
(backprop?) you are using, it is impossible to diagnose the problem.
Perhaps the net is doing as well as might be expected for this data set.
Perhaps you have chosen a poor net topology or parameters.

One thing does jump out, however: if you are trying to produce a
continuous-valued output, you might be better off with linear output units
than with sigmoids.  Then you won't have to pre-scale your data to fit the
sigmoid's range, and you won't be getting distortion due to gratuitous
nonlinearities.

-- Scott

===========================================================================
Scott E. Fahlman			Internet:  sef+ at cs.cmu.edu
Senior Research Scientist		Phone:     412 268-2575
School of Computer Science              Fax:       412 681-5739
Carnegie Mellon University		Latitude:  40:26:33 N
5000 Forbes Avenue			Longitude: 79:56:48 W
Pittsburgh, PA 15213
===========================================================================