Abstract
A study has been made to assess the level of predictive skill associated with the application of a simple analog methodology to long-range temperature prediction over the continental United States in the period 1948–78. This approach relies solely on the pattern correlation statistic to select analogs from which monthly and seasonal temperature category forecasts for 68 climatic divisions (CD's) are subsequently, derived. Numerous analog model trials were attempted, employing various combinations of predictor type and domain, forecast period, and forecast lead time. Predictor types include North Pacific sea surface temperature (SST), 700 mb height and 1000–700 mb thickness. The mean percent correct statistic was used to assess the spatial and temporal variations of skill for each analog model trial, as well as for persistence forecasts.
Principal conclusions include:
1) Overall mean percent correct scores for both monthly and seasonal analog models (using three categories) were, for the most part, slightly better than random chance and occasionally better than persistence. Highest overall scores were 45% correct for February forecasts using January 700 mb heights, and 40% correct for winter forecasts using November and fall SST. Counts of significant local skill exceeded chance expectation for many analog models tested.
2) Monthly analog models generally performed best during the period January–June, outscoring persistence and chance in many instances.
3) Seasonal analog models did best for the winter and summer seasons. Winter forecasts were most successful using Pacific SST, while similar results were obtained for summer, using SST or 1000–700 mb thickness. Seasonal analog models also performed well for spring, relative to random chance and persistence, particularly those using 700 mb heights. Thickness models using a forecast lag of one season appeared to be the best overall, with some combinations of domain and lag beating persistence for each season.
