Evaluating the Limits of Predictability in the ECMWF

Despite steady and significant progress in the skill of global numerical weather prediction, "bust" or "dropout" events occur when forecast errors grow much larger than the mean performance of the model. Researchers at the European Center for Medium Range Weather Forecasts (ECMWF) investigated errors for day-6 predictions over Europe (Rodwell et al. 2013) utilizing the ERA-interim system as a fixed model. Their study identified 584 bust cases in a 22-year period in which the day-6 forecast root-mean-square error (RMSE) of 500 hPa heights exceeded 60m over Europe, and the anomaly correlation coefficient (ACC) dropped below 40 percent.

This study expands the examination of forecast busts to a global perspective. Forecasts of large-scale features around the globe are decomposed through an EOF analysis of 500 hPa heights. The PCA scores from these EOFs are then plotted in a time-lagged chart, based on Barnston et al. (2012). From this visualization, forecast trends can be easily diagnosed. A limit of predictability for each EOF is defined using a threshold RMSE between the forecast PCA score and verifying PCA score, which gives insight into the potential utility of the EOF in extended range prediction within the model, as well as its contribution to the overall RMSE. The definition of a bust is then extended to the cases with a much lower limit of predictability than is observed for a given EOF. Each bust case is clustered by the EOF with the highest error and the shortest limit of predictability.

Using Rossby wave tracing techniques, error growth is diagnosed within each cluster. It is found that distinct wave trains are associated with each cluster. The end of the wave train projects strongest onto the EOF in question. Error growth originates near the beginning of the wave train. It is proposed that the origins of error growth are linked to the upscale influence of convective forcing near the entrance to the respective waveguide, and the transport of errors downstream is made more efficient through an extension of the waveguide.