Multi-model calibration and combination of tropical seasonal sea surface temperature forecasts

Different combination methods based on multiple linear regression are explored to identify the conditions that lead to an improvement of seasonal forecast quality when individual operational dynamical systems and a statistical–empirical system are combined. A calibration of the post-processed output is included. The combination methods have been used to merge the ECMWF System 4, the NCEP CFSv2, the Météo-France System 3, and a simple statistical model based on SST lagged regression. The forecast quality was assessed from a deterministic and probabilistic point of view. SSTs averaged over three different tropical regions have been considered: the Niño3.4, the Subtropical Northern Atlantic and Western Tropical Indian SST indices. The forecast quality of these combinations is compared to the forecast quality of a simple multi-model (SMM) where all single models are equally weighted. The results show a large range of behaviours depending on the start date, target month and the index considered. Outperforming the SMM predictions is a difficult task for linear combination methods with the samples currently available in an operational context. The difficulty in the robust estimation of the weights due to the small samples available is one of the reasons that limit the potential benefit of the combination methods that assign unequal weights. However, these combination methods showed the capability to improve the forecast reliability and accuracy in a large proportion of cases. For example, the Forecast Assimilation method proved to be competitive against the SMM while the other combination methods outperformed the SMM when only a small number of forecast systems have skill. Therefore, the weighting does not outperform the SMM when the SMM is very skilful, but it reduces the risk of low skill situations that are found when several single forecast systems have a low skill.