Key points are not available for this paper at this time.
The theoretical framework developed by Barnett and Preisendorfer (1978) for multifield analog prediction of United States seasonal temperatures has been modified and expanded to accommodate the use of composites of analogs and antianalogs to form predictions. Major changes have also been made both in predictor data and in the way it is processed, although the general strategy of Barnett and Preisendorfer served as a guide in this regard. Cross‐validation tests on a 35‐year record suggest that substantial gains in winter forecast skill have been achieved through both the previously mentioned architectural changes and several predictor data set changes. The latter include the use of a different El Niño/Southern Oscillation index and United States surface temperature data but not precipitation data. It was found that significant model skill depends most on these two data sets, along with well‐filtered 700‐mbar heights, and depends least on sea surface temperatures. Considerable skill was found over the eastern half and the north‐central portion of the United States. Forecasts were found to be effectively independent of and to outperform those of persistence and were comparable in skill to official forecasts. In a quasi‐operational test most of the system's skill was reproduced, even under very disadvantageous circumstances. Because of all these factors, the mixed analog and antianalog prediction system has been adopted as a major input for operational use by official forecasters. Development of models for other seasons will be described in a subsequent paper.
Livezey et al. (Tue,) studied this question.