Reduction of model systematic error by statistical correction for dynamical seasonal predictions

Singular value decomposition analysis (SVDA) is used to analyze an ensemble of three 34-yr general circulation model (GCM) simulations forced with obs erved sea surface temperature. It is demonstrated low statistical postproce ssing based on the lending SVDA modes of simulated and observed fields, pri marily precipitation, can be applied to improve the skill of the simulation . For a given limited prediction region, the GCM has the potential to nonli nearly transform the SST information from around the globe and produce a dy namic solution over the region that can be statistically corrected to accou nt for such features as systematic shifts in the location of anomaly dipole s. This paper does not address the separate question of whether the current generation of GCMs contain information above that which could be extracted using linear statistical relationships with SST. For precipitation, examples are drawn from skillful tropical regions, as we ll as the moderate-to-low skill Pacific-North American and North Atlantic-E uropean regions. Skill averaged across the analysis domain, as measured by the mean anomaly correlation, is notably improved by the statistical postpr ocessing in almost all situations where there is at least some real skill i n the raw model fields. Postprocessing based on leading canonical correlati on analysis (CCA) modes has been compared to postprocessing based on leadin g SVDA modes. The two methods show small differences, but neither one of th e methods can be claimed to do better than the other. A third method, which is based on the leading empirical orthogonal functions of the simulations, has been tested on examples of tropical rainfall where it is shown to also be successful, but with skill generally a little below that based on SVDA or CCA modes. The statistical postprocessing appears to have the greatest potential to im prove skill for a variable like precipitation, which can have particularly strong anomaly gradients. Application of the postprocessing to large-scale atmospheric fields of 500-hPa geopotential height and sea level pressure pr oduced smaller skill improvements relative to the skill of the raw model ou tput.