univariate time series prediction with neural network

hi all, I want to use neural network for predicting a univariate time series. I have a series of 1000 points, I want to use a sliding window (the size of my window is 35 points) to predict next 5 points. I am beginner in neural network so I dont know how to choose my inputs and outputs ( since I have to enter each 35 points for a response of 5 points over all my 1000 points I imagine! ). plz could you help me by giving a detailed programm? I looked at time delay neural network, but I dont know how to choose the number of delays and the number of hidden layers.

thanks

Answer 1

[Greg Heath]

TIMEDELAYNET predicts future outputs from present and past inputs only

NARNET predicts future outputs from past outputs only

NARXNET predicts future outputs from past outputs in addition to present and past inputs.

Choose the delays from the significant values of the input/output crosscorrelation function at NONEGATIVE lags and output autocorrelation function at POSITIVE lags.

Choose the number of hidden nodes, H, by trial and error. The default is H = 10

[x,t] = simplenarx_dataset;

X = zscore(cell2mat(x));

T = zscore(cell2mat(t));

[ I N ] = size(X)

[ O N ] = size(T)

crosscorrXT = nncorr(X,T,N-1);

autocorrT = nncorr(T,T,N-1);

crosscorrXT(1:N-1) = []; % Delete negative delays

autocorrT(1:N-1) = [];

sigthresh95 = 0.21 % Significance threshold

sigcrossind = crosscorrXT( crosscorrXT >= sigthresh95 )

sigautoind = autocorrT( autocorrT >= sigthresh95 )

inputdelays = sigcrossind(sigcrossind <= 35)

feedbackdelays = sigautoind(sigautoind <= 35)

feedbackdelays(1)=[] % Delete zero delay

If your N = 1000 series is relatively stationary (i.e., the local means and variances are relatively constant) you should be able to reduce the window size below 35 (15?).

Hope this helps.

Thank you for formally accepting my answer.

Greg

[Marina]

Hello Greg and thanks for your answer, I am quite beginner in neural network .. So could you give me plz an example how to code this by NAR not by NARX ( NAR seems more adequate for my problem) And how can I perform mutli step prediction ahead .. I tested stationnarity in my series, and I have find that it is stationnary just in 35 points

Answer 2

[Greg Heath]

1. Search the website documentation for sample code and/or demo.

2. Search the command line documentation for sample code.

help narnet

doc narnet

3+4. Search Answers and the Newsgroup using

nar

narnet

5. Try some designs.

6. If you have problems, post the relevant code and complete error messages.

Have fun,

Greg

Answer 3

[Marina]

Hey Greg and thaks again.

Apparently my data are not stationnary so I can't catch the suitable d. form the auctocorrelation function . I made differnation in order to stationnarize them. But how can I make prediction of the initial data from the stationnarized data by neural network

Thank you

Answer 4

[Greg Heath]

If you differenced the original series to reduce nonstationarity, did you check the new series for stationarity? You may have to difference again.

if d(j) = o(j)-o(j-1) then

o(j) = o(j-1) + d(j)

Hope this helps.

Greg