Empirical Correction of the NCEP Global Forecast System

Abstract

Abstract This paper examines the extent to which an empirical correction method can improve forecasts of the National Centers for Environmental Prediction (NCEP) operational Global Forecast System. The empirical correction is based on adding a forcing term to the prognostic equations equal to the negative of the climatological tendency errors. The tendency errors are estimated by a least squares method using 6-, 12-, 18-, and 24-h forecast errors. Tests on independent verification data show that the empirical correction significantly reduces temperature biases nearly everywhere at all lead times up to at least 5 days but does not significantly reduce biases in forecast winds and humidity. Decomposing mean-square error into bias and random components reveals that the reduction in total mean-square error arises solely from reduction in bias. Interestingly, the empirical correction increases the random error slightly, but this increase is argued to be an artifact of the change in variance in the forecasts. The empirical correction also is found to reduce the bias more than traditional “after the fact” corrections. The latter result might be a consequence of the very different sample sizes available for estimation, but this difference in sample size is unavoidable in operational situations in which limited calibration data are available for a given forecast model.