How to use Automatic Relevance Determination (ARD) to determine the window-size in time series forecasting

Hello,

I want to do a one step ahead prediction with my time series using neural networks. I have read some where that using ARD we can determine the appropriate size of windows for creating the training set. In other words we can determine the appropriate number of inputs for the neural network. Could you please help me with that? I have a very limited knowledge of ARD and found this file : http://www.mathworks.com/matlabcentral/fileexchange/2654-netlab/content/demard.m However I am not sure how to use that for my case.

Thank you Shadan

Answer 1

[Greg Heath]

I am not familiar with ARD. I just use the significant lags in the input/target cross-correlation function and the target autocorrelation function.

It would be more accurate to use partial correlations with a stepwise approach that takes into consideration lags that were used in the previous step (Similar to stepwise fitting of a linear model using the function STEPWISEFIT). However, I have not tried this.

[Shadan]

Thank you for giving me alternative solutions. But could you please give me more details and directions? I'm new to this field.

[Greg Heath]

If you have the Econometrics toolbox(which I do not), you could use autocorr, crosscorr and parcorr

The first two should be relatively straightforward. I cannot help you with the latter.

You can find my suboptimal approach by searching

 greg nncorr

in both the NEWSGROUP and ANSWERS. The most recent post are the most reliable because nncorr is buggy.