What is the input to layer recurrent network for time-series prediction task?

In all the examples of `layrecnet` I've seen so far the task is to predict an output of a function based on an output from some other function. Like here: Design Layer-Recurrent Neural Networks. In my case I have only one-dimensional time-series data and I need to predict future values of it. How do I train layrecnet for that purpose?

I managed to do it with NARX network by supplying my y values for both inputs and targets during training. Somehow I can't do it with layrecnet and I started to get confused. Some intuition tells me that my input should be my data points and targets should be the same but shifted by 1 to the left. Am I even going the right direction? I'd appreciate examples.

Answer 1

[Greg Heath]

For documentation on the self prediction of a single series use NARNET

 help narnet

 doc narnet

For examples earch both NEWSGROUP and ANSWERS using

 greg narnet

Hope this helps.

Thank you for formally accepting my answer

Greg

PS Note that the autocorrelation function is used to determine which subset of the statistically significant lags to use.

[Pawel Gajewski]

Thank you, but does this mean that what I'm doing with NARX and Elman networks makes no sense in a single time series? I actually wanted to compare those two with NAR.

[Greg Heath]

Yes.

NARXNET and ELMANNET have separate input and target series, whereas NARNET has the same series for input and target.

NARNET and TIMEDELAYNET are special cases of NARXNET.

ELMANNET differs in that delayed hidden layer signals instead of delayed output signals are fed back to the input.

Hope this helps.

Greg

P.S. It seems to me that NARXNET should be able to duplicate any I/O relationship created by ELMANNET. However, I do not have a proof.

Perhaps a reader can elucidate.