Mixed analog/persistence prediction of seasonal mean temperatures for the USA

Abstract The operational analog prediction system used as input to official USA seasonal temperature category forecasts was modified to produce anomaly forecasts instead of most probable class forecasts. These modified forecasts led to more conservative predictions, but were more reliable for above‐and below‐normal classes. They were subsequently combined with persistence forecasts in a variety of objective ways in the hope of producing forecasts more skilful than either input. Several versions of two basic mixing approaches, bivariate linear regression and persistence plus analog‐predicted change, were tried. All combined schemes were, overall, more skilful than either of the component schemes. Bivariate linear regression slightly outperformed persistence plus analog‐predicted trend. The former was most skilful when use of the regression forecast was limited to cases where the regression coefficient for the analog term was positive and significance tests were passed at least to a minimal extent in a cross‐validated examination of the regression equation.