Multiple NARX problem with correlation in train and test sets

Hi all! I have next problem. When I build multiply NARX network, I control regression parameter (first of all - correlation coef. in train and test sets during network building). After that, using internal network indices (I mean tr.trainInd,tr.valInd,tr.testInd)I build own datasets. And then I calculate correlation of train and test sets using corrcoef. The problem is next: these correlation coefficients strongly NOT EQUAL! Why?