Seasonal predictability of tropical rainfall: probabilistic formulation and validation

Two idealized seasonal forecast experiments are performed by prescribing monthly observed SSTs to atmospheric GCMs. The first one uses 3 different models, each with 9 individual forecasts (PROVOST experiment). The second one uses an improved version of one of the 3 models and larger ensembles consisting of 120 members. Both experiments show that forecast scores are maximum in the tropics during winter and during summer. The relatively high correlations in the tropics (0.4 to 0.7) imply, however, that the forecasts explain less than 50% of the variance of the observations. The raw probabilistic forecasts obtained by the empirical probability distribution of the forecast members exhibit very little skill, when evaluated by a euclidian distance versus the climatological forecast. The lack of reliability can be partly corrected by a simple statistical adaptation. Moreover, when the skill is evaluated by an economical value in a cost/loss approach, the model forecasts are more efficient than the climatological forecast. A more realistic evaluation of the probabilistic skill is obtained by replacing observed by statistically predicted SSTs. A simple but efficient method is used, which lets each member of the ensemble develop its own SST anomalies. Although lower, skill is significant in the tropics.