Seasonal forecasting at ECMWF

The main scientific basis for seasonal prediction is that the lower boundary conditions, such as SST or soil moisture, may have a longer memory than the atmosphere, and hence be at least partly predictable. The atmosphere is a chaotic system, and consequently, its response to the forcing of the lower boundary conditions is not deterministic; in other words, the atmosphere responds to the external forcing, changing its probability density function (PDF). Hence, seasonal forecasting is a probabilistic problem, and consists in quantifying how the atmospheric PDF at a given time differs from that of climatology.

Any seasonal forecasting system will have three components: a) a way of predicting the lower boundary forcing, b) a way of quantifying the atmospheric response, and c) a way of estimating the atmospheric PDF and its departure from climatology. The seasonal forecasting at ECMWF is a "one tier approach": a coupled ocean-atmosphere model and an ocean data assimilation system provide a tool to predict the lower boundary forcing (SST and soil conditions) as well as a method to quantify the atmospheric response. The atmospheric PDF is sampled using an ensemble system, in which a number of coupled integrations (about 30 per month) are carried out using perturbed initial conditions. Probabilistic tests that compare the estimated PDF at a given time with the climatological PDF are used to issue the seasonal forecast.

The skill of a probalistic forecast system is not easy to measure, especially when the number of verification times is scarce, as is the case of predictions made at seasonal intervals. However, some verifications using past cases can give some guidelines about the quality of the forecasts. In particular, the major El Nino event of 1997-1998 provided useful test for the system.