Predictable Anomaly Patterns and the Forecast Skill of Northern Hemisphere Wintertime 500-mb Height Fields

Abstract

Techniques for identification of well-predicted spatial patterns innumerical weather prediction model output are outlined and appliedto a 14-winter set of Northern Hemisphere 500-mb geopotential heightanalyses and 1-10-day forecasts produced by the ECMWF operationalmodel. Three approaches are investigated: canonical correlation analysis(CCA), singular value decomposition analysis, and predictable componentanalysis, the products of which are related to the optimization offorecast-analysis correlation, covariance, and rms error, respectively.In confirmation of earlier results, the most predictable anomalypattern identified by all three methods is found to be similar tothe leading empirical orthogonal function of the analyzed 500-mbheight anomaly field, which is dominated by the Pacific-North Americanpattern. The time series of forecast and verifying analysis projectionsonto the leading pattern have temporal correlations of at least 0.75at all forecast intervals out to 10 days and greater than 0.85 for5-day averages of 6-10-day forecasts and analyses. The leading patterndisplays strong temporal persistence and is prominent on the interannualtimescale. CCA is found to be the most desirable technique for identificationof such patterns. When CCA is applied to the first seven winters'data (as a dependent sample), the amplitude of the leading patternis well predicted in either polarity and the skill of the full forecastfield is shown to increase as the amplitude of the leading patternincreases, regardless of the polarity. However, when the analyzedand predicted fields from the second seven winters of the dataset(an independent sample) are projected onto the patterns derived fromthe first seven winters, the skill of the full forecast field doesnot appear to be well related to the amplitude of the leading predictablepattern. Slight decreases in rms errors were achieved by statisticallycorrecting the independent data, bur only at the expense of a considerabledamping of forecast amplitude. It is concluded that continuing modelimprovements make such approaches to skill prediction and statisticalcorrection of little value in an operational setting.