Ensemble forecasting has been operational at NCEP (formerly the Nation
al Meteorological Center) since December 1992. In March 1994, more ens
emble forecast members were added. In the new configuration, 17 foreca
sts with the NCEP global model are run every day, our to 16-day lead t
ime. Beyond the 3 control forecasts (a T126 and a T62 resolution contr
ol at 0000 UTC and a T126 control at 1200 UTC), 14 perturbed forecasts
are made at the reduced T62 resolution. Global products from the ense
mble forecasts are available from NCEP via anonymous FTP. The initial
perturbation vectors are derived from seven independent breeding cycle
s, where the fast-growing nonlinear perturbations grow freely, apart f
rom the periodic rescaling that keeps their magnitude compatible with
the estimated uncertainty within the control analysis. The breeding pr
ocess is an integral part of the extended-range forecasts, and the gen
eration of the initial perturbations for the ensemble is done at no co
mputational cost beyond that of running the forecasts. A number of gra
phical forecast products derived from the ensemble are available to th
e users, including forecasters at the Hydrometeorological Prediction C
enter and the Climate Prediction Center of NCEP. The products include
the ensemble and cluster means, standard deviations, and probabilities
of different events. One of the most widely used products is the ''sp
aghetti'' diagram where a single map contains all 17 ensemble forecast
s, as depicted by a selected contour level of a field, for example, 55
20 m at 500-hPa height or 50 m s(-1) windspeed at the jet level. With
the aid of the above graphical displays and also by objective verifica
tion, the authors have established that the ensemble can provide valua
ble information for both the short and extended range. In particular,
the ensemble can indicate potential problems with the high-resolution
control that occurs on rare occasions in the short range. Most of the
time, the ''cloud'' of the ensemble encompasses the verification, thus
providing a set of alternate possible scenarios beyond that of the co
ntrol. Moreover, the ensemble provides a more consistent outlook for t
he future. While consecutive control forecasts verifying on a particul
ar date may often display large ''jumps'' from one day to the next, th
e ensemble changes much less, and its envelope of solutions typically
remains unchanged. In addition, the ensemble extends the practical lim
it of weather forecasting by about a day. For example, significant new
weather systems (blocking, extratropical cyclones, etc.) are usually
detected by some ensemble members a day earlier than by the high-resol
ution control. Similarly, the ensemble mean improves forecast skill by
a day or more in the medium to extended range, with respect to the sk
ill of the control. The ensemble is also useful in pointing out areas
and times where the spread within the ensemble is high and consequentl
y low skill can be expected and, conversely, those cases in which fore
casters can make a confident extended-range forecast because the low e
nsemble spread indicates high predictability. Another possible applica
tion of the ensemble is identifying potential model errors. A case of
low ensemble spread with all forecasts verifying poorly may be an indi
cation of model bias. The advantage of the ensemble approach is that i
t can potentially indicate a systematic bias even for a single case, w
hile studies using only a control forecast need to average many cases.