Seasonal statistical–dynamical prediction of the North Atlantic Oscillation by probabilistic post-processing and its evaluation

Abstract

Dynamical models of various centres have shown in recent years seasonal prediction skill of the North Atlantic Oscillation (NAO). By filtering the ensemble members on the basis of statistical predictors, known as subsampling, it is possible to achieve even higher prediction skill. In this study the aim is to design a generalisation of the subsampling approach and establish it as a post-processing procedure. Instead of selecting discrete ensemble members for each year, as the subsampling approach does, the distributions of ensembles and statistical predictors are combined to create a probabilistic prediction of the winter NAO. By comparing the combined statistical–dynamical prediction with the predictions of its single components, it can be shown that it achieves similar results to the statistical prediction. At the same time it can be shown that, unlike the statistical prediction, the combined prediction has fewer years where it performs worse than the dynamical prediction. By applying the gained distributions to other meteorological variables, like geopotential height, precipitation and surface temperature, it can be shown that evaluating prediction skill depends highly on the chosen metric. Besides the common anomaly correlation (ACC) this study also presents scores based on the Earth mover's distance (EMD) and the integrated quadratic distance (IQD), which are designed to evaluate skills of probabilistic predictions. It shows that by evaluating the predictions for each year separately compared to applying a metric to all years at the same time, like correlation-based metrics, leads to different interpretations of the analysis.