Abstract
Technique for identification of well-predicted spatial patterns in numerical weather prediction model output are outlined and applied to a 14-winter set of Northern Hemisphere 500-mb geopotential height analyses and 1–10-day forecasts produced by the ECMWF operational model. Three approaches are investigated: canonical correlation analysis (CCA), singular value decomposition analysis, and predictable component analysis, the products of which are related to the optimization of forecast-analysis correlation, covariance, and rms error, respectively. In confirmation of earlier results, the most predictable anomaly pattern identified by all three methods is found to be similar to the leading empirical orthogonal function of the analyzed 500-mb height anomaly field, which is dominated by the Pacific-North American pattern. The time series of forecast and verifying analysis projections onto the leading pattern have temporal correlations of at least 0.75 at all forecast intervals out to 10 days and greater than 0.85 for 5-day averages of 6–10-day forecasts and analyses. The leading pattern displays strong temporal persistence and is prominent on the interannual timescale. CCA is found to be the most desirable technique for identification of such patterns.
When CCA is applied to the first seven winters' data (as a dependent sample), the amplitude of the leading pattern is well predicted in either polarity and the skill of the full forecast field is shown to increase as the amplitude of the leading pattern increases, regardless of the polarity. However, when the analyzed and predicted fields from the second seven winters of the dataset (an independent sample) are projected onto the patterns derived from the first seven winters, the skill of the full forecast field does not appear to be well related to the amplitude of the leading predictable pattern. Slight decreases in rms error were achieved by statistically correcting the independent data, but only at the expense of a considerable damping of forecast amplitude. It is concluded that continuing model improvements make such approaches to skill prediction and statistical correction of little value in an operational setting.
