In real time since 1990, the National Meteorological Center (NMC) has
been running a system to predict the forecast skill of the medium-rang
e forecasts produced by the NMC global spectral model. The predictors
used are the agreement of an ensemble consisting of operational foreca
sts from various centers, the persistence in the forecast, and the amp
litude of the anomalies. These predictors are used in a stepwise regre
ssion scheme, with the last 60 days used as training period, and the r
egional anomaly correlation of the 0000 UTC NMC global forecast is pre
dicted from days 1 to 6. By far the most important predictor of skill
is the agreement between the NMC global forecast started at 0000 UTC,
out to 6 days, and four other 12-h ''older'' forecasts (Japan Meteorol
ogical Agency, United Kingdom Meteorological Office, and the European
Centre for Medium-Range Weather Forecasts, as well as the average of t
he NMC forecast at 0000 UTC with the previous day's forecast). The oth
er predictors have been selected to add to the predictive capability o
f the agreement alone, and together they quantify the factors that for
ecasters use subjectively when evaluating the available forecasts. The
se predictions are available to NMC forecasters on workstations and to
outside users through the Internet. The predictive ability of this sy
stem compares favorably with recent theoretical and experimental studi
es. The correlation between predicted and verifying forecast skill see
ms to be best in regions where forecast skill varies significantly. Th
e seasonal variation in predicting the skill is small except in the Tr
opics. The overall performance shows that these predictors include eno
ugh information about forecast skill to justify further development of
skill predictions based on larger forecast ensembles and on more soph
isticated statistical techniques.