Incorrect prediction using svmtrain/svmpredict

Dear all,

I've been using the libsvm package to predict a timeseries (regression analysis using SVM). The code I've used is below:

***

A = csvread('norm agv BDI.csv');

% generate sin function points with period of 5 years (60 months)

A(:,3) = sin(A(:,2)*2*pi/60);

% normalize values from [-1,1] to [0,1]

A(:,3) = (A(:,3)+1)./2;

plot(A(1:360,2),A(1:360,3),'or');

hold on

% denoise data:

xd = wden(A(:,3)','heursure','s','one',3,'db3')';

plot(A(1:360,2),xd(1:360),'-b','LineWidth',5);

trainfeatures = [A(:,2)];

model = svmtrain(xd(1:336), trainfeatures(1:336,:), '-s 4 -t 2 -c 5.5 -e 0.02');

% options are:

% -s SVM type: 3 - epsilon SVR, 4 - nuSVR
% -t kernel type: 2 radial basis function: exp(-gamma*|u-v|^2)
% -c cost (value of parameter C)
% -e epsilon: tolerance of termination criterion

[predicted_label] = svmpredict(xd, trainfeatures, model);

[A(:,2), predicted_label, xd]

% now overlay the predicted model

plot(A(:,2),predicted_label,'-g','LineWidth',2);

The first two columns of vector A are read from a file which starts like this:

31533.00,0,-0.932541478

31564.00,1,-0.935964688

31594.00,2,-0.949384932

31625.00,3,-0.946143266

31656.00,4,-0.916051242

31686.00,5,-0.908048373

31717.00,6,-0.911407848

31747.00,7,-0.924103013

31778.00,8,-0.905556669

31809.00,9,-0.895828641

...

First column is day number (never used), second column is month, last column is timeseries value for the corresponding month.Ignore for now the denoising bit, that is there for another timeseries that requires it, for now I am just testing the code using a sine function rescaled from 0 to 1.

I am using all the 360-month data, except for the last 24 months, to generate an SVM model using svmtrain. Then I am feeding this model in svmpredict to get everything (including the last 24 months). As you can see in the figure below, I am getting excellent agreement on the training set (red - original data, blue - denoised, green - predicted), but as I look at the last 24 months I get a flat green line. I tried playing around with the C and epsilon parameters but got same results regardless of their values.

Any idea what might be wrong here?

Cristian