Seasonal Prediction of North American Surface Air Temperatures Using Space–Time Principal Components

Abstract

The statistical model proposed by Vautard et al. is applied to the seasonal prediction of surface air temperatures over North America (Canada and the United States). This model is based on sea surface temperature predictors filtered by multichannel singular spectrum analysis (MSSA), which is equivalent here to a nonseasonal version of extended EOF analysis. Several versions of the MSSA model are proposed. The most successful one is based on a two-step procedure consisting in a prior prediction of filtered sea surface temperatures followed by a predictand specification stage. The MSSA model is compared with the recent prediction technique based on canonical correlation analysis (CCA). The former model turns out, in this application, to be more skillful in most seasons than the latter. The differences are, however, marginal. The authors argue that these differences are due to the nonseasonal nature of the MSSA model and to overfitting problems inherent to CCA. Another advantage of the MSSA model relative to CCA is the possibility of easily transforming deterministic continuous forecasts into probabilistic categorical forecasts. The geographical distribution of prediction skill across North America is studied. Canada turns out to be the country where skill is most significant. During winter, high skill values are also found over the southeastern United States.