On the Improvement of Seasonal Predictability from Multi Model Ensembles

Meteorological Research Institute (METRI) multi model ensemble system for seasonal prediction has been established with four different climate models. Multi model hindcast ensemble experiment from 1979 to 2002 by using the same initial and boundary condition is performed for winter season. To evaluate seasonal predictability of linear multi model ensemble forecasts compared with each participating model, we have used quantitative verification methods as mean square error and skill score using climate as a control forecast. Geopotential height at the 500 hPa level, temperature at the 850 hPa level, and precipitation are used as verification variables with both time and space domains. NCEP/NCAR reanalysis and CMAP precipitation data are considered as observation data and climatology is defined as the average from 1979 to 2002. Cross-validation is applied to the verification process. The single models successfully simulated the northern hemispheric winter climate features; however, the seasonal predictability of each model was low in the midto high-latitudes. The interannual variability of mean square error (MSE) from multi model ensemble is found to be dependent on MSE of single models and the improvement of skill score from multi model ensembles (regression-improved and superensemble) are primarily attributed by the reduction of conditional bias in the model.