Seasonal Temperature Forecasts as Products of Antecedent Linear and Spatial Temperature Arrays

Abstract

Our objective is to evaluate the potential for extracting the maximum information contained in antecedent temperature patterns that operationally could be used in formulating winter seasonal forecasts in the United States. In particular, examination of the predictability of winter temperatures given autumn temperatures is made using derived contingency tables, discriminant equations of antecedent principal components, and canonical correlation analysis. Contingency tables were constructed based on tercile classifications of a seventy-five year dependent record (1895?1969). Testing of an independent data period (1970?78) using these tables produced winter forecasts with no appreciable skill in the aggregate (?0.04). Discriminant analysis deprived linear combinations of the five principal components of the antecedent seasonal (autumn) temperatures to distinguish between specific terciles of the predictand season (winter). Despite encouraging results for the dependent period, forecast skill for the independent test period achieved no significant score (?0.04). Unfortunately, both of these forms of analysis suffer imposed spatial limitations which restrict the scope of our investigation. Canonical correlation analysis is capable of relating the total spatial variance of fall temperatures to that of the winter temperatures for the entire United States. In this study, the technique was used to isolate seasonal patterns in winter temperature data that are correlated in time with fall temperature patterns for the same region. Summation of the first 20 canonical variate pairs suggests that autumn and winter temperatures over the continental United States are not closely related to one another.