An attempt was made to predict categories of cold season temperature for 19 cities in the central and eastern United States. Classification equations were developed by using discriminant analysis based on 23 years of historical data (1949--1971) in which the discriminating variables (predictors) are principal components of the mean 700-mbar heights during November over the northern hemisphere; sea surface temperatures of the eastern North Pacific, eastern tropical Pacific, and North Atlantic oceans; and a southern oscillation index for the fall season. Separate sets of equations were developed for varying duration of the prediction, from 1 to 4 months, and for two to three categories of mean temperature. An independent sample of years was reserved for testing the reliability of the equations. 'Forecasts' made by each set of equations were compared with observed categories of mean temperature for five recent winters (1972--1976), not included in the dependent sample. The sets of equations derived for discriminating between above- and below-normal mean temperatures for the 4-month season (December--March) and between three groups for the 3-month winter season (December--February) performed significantly better than chance expectation. Values of 63% and 45%, respectively, were correct. Additionally, when the persistence forecast category was the same as the forecast category of our model for the December--March period in the two-group discrimination, the prediction category was correct 86% of the time over the five test winters. This situation occurred, on the average, for about half the cities in each year. These results indicate that objective methods, using statistical relationships, can be useful for winter temperature predictions, but considerable improvement is still needed.