Global Sea Surface Temperature Forecasts Using an Improved Multimodel Approach

With the availability of hindcasts or real-time forecasts from a number of coupled climate models, multimodel ensemble forecasting systems have gained popularity in recent years. However, many models share similar physics or modeling processes, which may lead to similar (or strongly correlated) forecasts. Assigning equal weights to each model in space and time may result in a biased forecast with narrower confidence limits than is appropriate. Although methods for combining forecasts that take into consideration differences in model accuracy over space and time exist, they suffer from a lack of consideration of the intermodel dependence that may exist. This study proposes an approach that considers the dependence among models while combining multimodel ensemble forecast. The approach is evaluated by combining sea surface temperature (SST) forecasts from five climate models for the period 1960-2005. The variable of interest, the monthly global sea surface temperature anomalies (SSTA) at a 5 degrees x 5 degrees latitude-longitude grid, is predicted three months in advance using the proposed algorithm. Results indicate that the proposed approach offers consistent and significant improvements for all the seasons over the majority of grid points compared to the case in which the dependence among the models is ignored. Consequently, the proposed approach of combining multiple models, taking into account the interdependence that exists, provides an attractive strategy to develop improved SST forecasts.