Experimental Tropical SST Forecasts at IRI using Models of Different Complexity: Skill Evaluation and Diagnosis of Systematic Errors

David G. DeWitt

International Research Institute for Climate Prediction (IRI) Columbia University


A directly coupled atmosphere-ocean general circulation model (GCM) has been constructed from the ECHAM4.5 atmospheric GCM (AGCM) and the MOM3 ocean GCM (OGCM). The coupled model has been used to make hindcasts (and now forecasts) for all initial condition months in the 1982 to present period. Ocean initial conditions for the forecasts are obtained from a version of the Derber-Rosati ocean data assimilation (ODA) scheme. The forecasting system utilizes an ensemble of 7 members which differ only in the initial atmospheric state. Model forecast skill for SST in the central and eastern Tropical Pacific is found to be competitive for the 1 to 6 month lead period with other systems described in the literature. Forecast skill will be examined as a function of initial condition month as this is argued to be more relevant then the standard bulk ACC and RMSE indices which will also be shown and for which the model also performs competitively with other forecast systems. A brief examination of forecast skill will be made for the tropical Indian and Atlantic oceans as well. Skill in those regions is generally lower and more variable then that found in the tropical Pacific. Since the forecast system uses ensembles potential probabilistic forecast skill can be evaluated. This is done using the relative operating characteristic (ROC) metric.

A nearly universal systematic error in directly coupled models is an inability to correctly simulated the near Equatorial SST slope in the Atlantic Ocean. Diagnosis of the factors leading to the error in this coupled model will be explored.

Time permitting some brief discussion will be given to some other SST forecasting systems including a simple multi-model SST forecasting system, and non- OGCM based forecasting systems.