Prediction with Narxnet without future inputs

I'm a bit confused if prediction with Narxnet requires knowledge of future input. Documentation defines Narxnet as dependent only on past inputs and past outputs, but I cant seem to find a way of using narxnet without future inputs and unable to get any meaningful predictions with zero/nan input. How do I forecast a timeseries which is dependent on some inputs whose future values aren't available.

Answer 1

[Greg Heath]

You are correct. If you do not have future inputs you are in serious trouble. My approach is to design two additional NARNETS: One for the input and one for the output. Then you have 2 methods for n > N = length of original data.

1. Extend the closeloop output NARNET

2. Extend the closeloop input NARNET and use this in the original closeloop NARXNET.

3. Compare the answers and pray for divine guidance.

Hope this helps.

Thank you for formally accepting my fantastic answer

Greg

[cr]

Greg, Thanks! The description says "NARX (Nonlinear autoregressive with external input) networks can learn to predict one time series given past values of the same time series, the feedback input, and another time series, called the external or exogenous time series." So is this incorrect? (or did I interpret it wrong?)

If I'm predicting with Narnet, say predicting 10 points beyond the last data value, is this the right way:

netc2 = closeloop(net);
[xc,xic,aic,tc] = preparets(netc2,{},{},targetSeries);
yc2 = netc2(NaN(10,1),xic,aic);

In many of your answers you said 0:1 is an invalid delay although Matlab allows it; so, if removedelay() results in such a delay value do I need to use more delays in the architecture?

You thanked, but I've not formally accepted your fantastic answer yet :). Doing it now.

[Greg Heath]

CLARIFICATION:

0 is NOT A VALID feedback delay for CLOSELOOP configurations!

0 IS A VALID feedback delay for the OPEN LOOP configuration.

0 IS A VALID input delay for both configurations.

Hope this helps.