The Variability in Skill of 72-hour Global-Scale NMC Forecasts

Abstract

The variability in skill of NMC 72-h 500 mb forecasts during recent winter is examined. Root-mean-square error, anomaly correlation, and the Fisher z-transformation of the anomaly correlation are used as measures of skill. This latter score is appropriate for measuring variability because it produces nearly Gaussian distributions of scores. The annual mean skill of the forecasts improves throughout the period examined, but there is no trend in the variability of the z-transformed anomaly correlations. Thus model improvements do not seem to have improved forecast reliability. The temporal power spectrum of skill is red and the 1-day lag correlation of anomaly correlation scores is 0.60 during our sample period. Heights at 500 mb have a similar spectrum during the sampling period, so it may be that certain elements of the atmospheric flow can be used to predict the likely skill of a forecast. A few potential predictors are tested and some of these, e.g., the persistence of the flow and the spatial standard deviation of the predicted anomalies, are shown to predict 10 to 20% of the variance in skill.