Meanwhile, regional climate prediction skill at various time scales remains modest (IPCC 2001). Contrasting SRES ensemble simulations for seasonal rainfall over Southern Africa (forecasting a likely decrease) and the West African hinterland (poorly specified forecast) further suggest that variability and predictability do not necessarily go hand in hand. Models' knowledge base originally tuned to maximize performance over the Pacific region on interannual timescales (CLIVAR 1999) and relying on a subset of mostly oceanic and atmospheric predictors works satisfactorily when ENSO wields a dominant control over regional climate (even when interannual variability is highest: Southern Africa) but fails when the distribution of forcings is widespread (e.g. West Africa). Sometimes simple statistical methods outperform dynamical models constrained by poor initialization of regional soil moisture and lack of dynamically prescribed vegetation (Garric et al. 2002). Low local skill levels dominate in spite of an understandable urge to demonstrate the value of seasonal forecasts through more attractive scores at the aggregate level. For example, 'high degrees of skill' for the JAS period (IRI 2005) should be carefully interpreted in terms of scale-compatible applications (such as large watershed management), because any space-time downscaling will irremediably result in a loss of skill as suggested by Gong et al. (2003). The inability of dynamic models to correctly reproduce the succession of sub-grid scale convec-tive events severely limits their applicability in hydrology (Lebel et al. 2000) and even more so in smallholder agriculture.
Was this article helpful?