Regression trees, improve the test performance

Hi everyone,

I built a basic regression tree code for a complicated problem (27 inputs, one output, 630 timesteps)

rtree = RegressionTree.fit(InputsTr,Mn2tr,'Minparent',30);
view(rtree,'mode','graph');
Ynew = predict(rtree,Inputsts);

plot(Ynew)
hold on
plot(Mn2ts,'g')
hold off
err=Ynew-Mn2ts;
Msetst=mse(err)
percMse=Msetst/mean(Mn2ts)

I got a very good performance for the training set, but when I tried with new independent inputs (Inputsts), the performance dropped to very poor values.

Any suggestions about how to improve it? I have tried with pruning, making it simpler, but the improvement is very small (only because the resulting values are less noisy, but very far from the real targets)

Thank you