FORMULATION OF GAUSSIAN PROBABILITY FORECASTS BASED ON MODEL EXTENDED-RANGE INTEGRATIONS

A sample of 40, 44-day winter forecasts is used to investigate the pre dictability of 850 hPa temperature over Europe. These forecasts exhibi t a significant skill when averages of day 5 to day 14 and day 15 to d ay 44 are considered. This skill is, however, very close to that of th e trivial climatology forecast. A probability forecast is performed, u sing a gaussian density with the deterministic forecast for the mean, and the climatological standard deviation (SD). The rank probability s core (RPS) of such a forecast is better than, but again very close to, that of the probabilistic climatology forecast. The categorical forec ast is also studied as a limit case when the SD is zero. The RPS is mi nimal when using the conditional probabilities of the verification ana lyses, but the results are not widely improved when robust estimates a re used. The results could be widely improved if we used a suitable SD in our forecasts. However, the attempts to predict a priori the optim al SD lead to non-significant results. The best available probability forecast, in our local gaussian approach, uses, for the mean, the regr ession of the verification analyses by the model forecasts, and, for t he SD, a scaled climatological SD.