Neural Network Dynamic Time Series Analysis (NARX) with Multiple Series

I am trying to train a single NARX network using multiple series (each with their own set of inputs and target), but it seems that the the tool can only handle one time series.

This is not my application, but for example, say I wanted to train a single NARX network using time series data from 3 different securities/stocks. For each security, the inputs would be trading volume, high, low, etc. and the target would be the closing price. For each unit of time, I will have 3 of each input variable, and 3 targets. I'm not sure how to set this up as a time series NN in Matlab, and train a single net (as opposed to training a net for each series). Training individual nets for each series would loose some of the value with my multiple series

Is there a way for the NN toolbox to handle this type of training?

[Mei]

How did you make this work? Did you have a (3 x n_timesteps) cell array where each row represented each stock, and each cell being a 3 x 1 double, where each row represented the three input variables?

Answer 1

[Greg Heath]

I haven't looked in detail. However, I can't think of any reason why a narxnet cannot have multiple inputs and outputs.

The only drawback is that all of the input delays must be the same and all of the feedack delays must be the same.

Therefore, you should take a good look at the significant output autocorrelation peaks for positive lags and the significant input/output cross correlation peaks for nonnegative lags.

For x and t row vectors of the same length, N:

   lag0 = -(N-1) : (N-1);
   autocorrtt0  = nncorr(zscore(t,1),zscoret(t,1),N-1,'biased') ;
   crosscorrtx0 = nncorr(zscore(t,1),zscoret(x,1),N-1,'biased') ;

   lag = 0:N-1;
   autocorrtt   = autocorrtt0(N:end);
   crosscorrtx  = crosscorrtx0(N:end);

Hope this helps.

Thank you for formally accepting my answer.

Greg