A large experiment is performed to validate two predictors for the qua
lity of ECMWF forecasts over Western Europe. One predictor yields the
spread of the probability distribution for the error in the predicted
500 hPa geopotential height. It is determined by the trace of the cova
riance matrix for the geographically local forecast error. In addition
the spread for the 500 hPa vorticity error at a location near the Net
herlands is computed. The local covariance matrix, necessary for deter
mining both predictors, is computed for 607 days, using the tangent li
near and an adjoint version of a quasi-geostrophic 3-level model with
truncation T21. We assume linear error growth and the absence of model
errors. The forward reference orbit is obtained by interpolating actu
al ECMWF forecasts with the 3-level model. Small values of the predict
or imply small error growth, and therefore accurate forecasts. Large v
alues may or may not be associated with large actual forecast errors,
depending on whether the initial error strongly projects on the fastes
t growing modes. How the uncertainty in the structure of the initial e
rror influences the performance of the skill predictor is studied by c
onsidering three different covariance matrices for the initial error.
Validation of the predicted variance with the 2-day and 3-day ECMWF fo
recast error shows that for all initial covariance matrices, both pred
ictors provide significant information about the quality of the foreca
st. In case of small and large predicted variance, the probabilities f
or small and large prediction errors are 10% higher than the climatolo
gical probabilities. Projection of the observed forecast error onto th
e eigenvectors of the local covariance matrix indicates that a few eig
envectors already describe a large portion of the forecast error.