LONG-RANGE ATMOSPHERIC PREDICTABILITY USING SPACE-TIME PRINCIPAL COMPONENTS

The long-term predictability of 70-kPa geopotential heights is examine d by a series of hindcast experiments over a validation period of 40 y ears using empirical models. Only the North Atlantic sector is conside red. Significant skill is found up to lead times of one to two months for forecasts of time averages and of weather regime occurrence freque ncies. The empirical schemes produce forecasts of the conditional prob ability of occurrence of a predictand within its natural terciles. The se probabilistic forecasts are compared for two sets of predictors. Th e (spatial) principal components of the Atlantic large-scale Bow (S-PC s) and its space-time principal components (ST-PCs) obtained from mult ichannel sin gular spectrum analysis (MSSA). These latter predictors a chieve a good compromise between explained variance and predictability . In particular, the skill of a one-step model, where predictand's con ditional probabilities are obtained directly from an analog method, is compared with a two-step model, which first forecasts the ST-PCs and then specifies the predictand's conditional probabilities. The onestep model is systematically beaten by the ST-PC scheme for lead times bey ond 10 days. An attempt is made to explain why ST-PCs perform better t han S-PCs in the long run by applying the forecast schemes to a simple low-order chaotic dynamical system. The key factor seems to be that f or a dynamical system displaying low-frequency behavior and nonlinear spells of oscillations, the MSSA expansion gathers these phenomena int o a few leading ST-PCs. These ST-PCs are therefore good candidates to quantify the concept of atmospheric ''predictable components.''