Determining the Time series prediction

Hi all, according to simpleseries_dataset code in neural network there is a difference between it and NAREXNET. Is it in the coding or in the implementation of the function itself?

[Greg Heath]

I don't follow. Comparing

 help simpleseries_dataset

with

 help narxnet

All I see is

the use of PLOTRESPONSE instead PERF.

Please explain your comment in more detail.

Greg

[Lilya]

sorry for that. I trained the network with NARXNET the perf. doesn't close to 0 but the regression is almost close to the actual line. The attached file shows what I mean.

Thank you Dr. For all your help.

[Greg Heath]

I don't agree.

The high correlation between input and error is indicative of a poor fit.

This would become more evident if you posted

 plotperform

and

 plotregression

Hope this helps,

Greg

[Lilya]

Dr. heath, How can I determine the number of the forecast horizon in the NN code? Its depend on the ID or FD?

Thank you

[Greg Heath]

The maximum lag from both ID and FD.

[Lilya]

I got them from your answers I'm really thank you so much

Answer 1

[Greg Heath]

GOOD QUESTION!

My answer is TRIAL and ERROR

The advice I usually give for starting the process is

 1) Use divideblock datadivision. 
 2) First use the default 0.7/0.15/0.15
 3) Use the training data to estimate the
     a. significant target autocorrelation lags
     b. significant input-target crosscorrelation lags
 4) Use 2, 3 and corresponding plots for lags 0 to
    Ntrn/2 to guide a choice for ID and FD. 
 5) Determine the minimum number of hidden nodes for a 
specified (degree-of-freedom adjusted) training error rate 
e.g., NMSEtrna < 0.005 )
 6) If successful try decreasing Ntrn
 7) Using the smallest acceptable Ntrn for the openloop configuration, close the loop 
and investigate the closeloop configuration.

Hope this helps.

Greg

[Lilya]

Excuse me Dr I have another question Is it important to shuffling Inputs?

The final plot and performance are different when I shuffle data

[Greg Heath]

Recommendation #1 for TIMESERIES DATA is to use DIVIDEBLOCK in order TO NOT SHUFFLE THE DATA!