A Further Assessment of Winter Temperature Predictions Using Objective Methods

Abstract

Various statistical models were tested for their reliability in predicting mean winter temperatures in the eastern and central United States. This study was designed td extend, and improve upon, a previous study, which also used statistical methods to formulate and test models for forecasting winter temperatures. In the current study, principal component analysis was performed separately on various predictor fields, which were selected mainly on the basis of physical-statistical relationships proposed in the literature. These predictor fields included the mean monthly sea-surface-temperature (SST) distribution for November in the eastern North Pacific Ocean. in the eastern tropical Pacific Ocean and in the North Atlantic Ocean, as well as the November 700 mb heights for a portion of the Northern Hemisphere. The time-varying amplitudes. which are derived by applying principal component analysis, were the predictors used as input for screening multiple linear regression. A Southern Oscillation (SO) index for the fall season was used as an additional predictor. The predictands consisted of area-averaged winter temperatures for nine sub-areas of the region cast of the Rocky Mountains. Four sets were formulated. corresponding to four lengths of winter: December only, December?January, December?February and December?March. In one experiment all the potential predictors assembled (23) were used in turn with each set of predictands as input to the regression analysis. The dependent sample consisted of the period 1949?71 (23 years). The resulting four sets of prediction equations were then tested on an independent sample consisting of winters in the period 1972?77 (six years). Forecast and observed winter temperatures were put into categories for verification purposes. lime temperature classes were used. The main results of this testing showed that the forecasts of winter temperature categories improved as the length of the winter period used in the model increased. The mean percent correct for the December-only model was 20%, increasing to 56% for the December?March model. The expected mean percent by chance would be 33%. In a second experiment, various subsets of the full act of predictors assembled were selected in order to formulate various types of predictor models. The predictor models consisted of an SST-only model, a circulation-only model (i.e., 700 mb height components plus the 50 index), and an all-predictors model (circulation and SST predictors) having a reduced number of components. Each of these predictor acts was matched with the predictand set consisting of mean temperatures for the December?February winter period. The main results obtained by testing each of the sets of prediction equations on the independent sample showed that the SST-only model was superior to the others (56% correct) and persistence. The circulation-only model had 33% correct, while the all-predictor reduced component model had 30% correct. More importantly, the SST-only model was able to distinguish correctly between the mild winter of 1975?76 and the cold winters of 1976?77 and 1977?78. Supporting diagnostic work is also presented.