It looks like you are trying to predict a timeseries of 256 dimensional vectors beyond the known 300 examples.
1. Ordinarily, NARNET would be used. 2. However, it is only practical for timeseries vectors of small dimension (e.g., < 10). See, for example, the practice datasets obtained from
help nndatasets
doc nndatasets
3. So, I suggest you use the specifics of the problem origin to drastically reduce the dimensionality of the vectors.
4. Next would be to determine the significant autocorrelation and crosscorrelation lags among the remaining dimensions.
5. You probably should discuss the essence of the problem you are trying to solve with an expert in the field. It doesn't seem to me to be an appropriate way to solve any practical problem that I can think of.
Hope this helps.
Greg
Best Answer