Ensemble forecast products from both NCEP and ECMWF are evaluated. To study how much the forecast errors can be explained by the ensemble perturbations, we compare the structures of forecast errors and the first 10 individual ensemble perturbations. For different forecast lead times, the correlations between the forecast errors and the individual ensemble perturbations are computed. In addition, the optimally combined perturbations of the first 10 ensemble perturbations are computed and also compared with the forecast errors. The structures of the residual vectors between the optimally combined vectors and forecast errors are analyzed to identify those dominant structures which cannot be explained by the ensemble perturbations. The dependence of correlations on the number of ensembles is studied for different lead times by using data from both centers.

It is found that the performances of both ensemble forecast systems are similar globally in most cases, with NCEP ensembles having slightly advantages during the first 2 days in some cases. But after about 2 days, ECMWF ensembles have higher correlations in some cases. For longer lead times, the correlations from both systems are converging, indicating that both forecast errors and singular vectors are converging to the dominant Lyapunov vector. In tropic region, NCEP ensembles perform than better ECMEF's in most cases.