Statistical and dynamical long-range atmospheric forecasts: experimental comparison and hybridization

We perform a direct comparison between a statistical forecast model using space-time principal components as predictors and a series of experimental long-range (up to 44 days) dynamical forecasts performed at Météo-France using a simplified-physics version of the former french operational forecast model “Emeraude”. The comparison is made possible by forecasting the same upper air quantity, the monthly averaged 50 kPa geopotential height, at the same forecast dates and lead times, using the same skill measure. A disappointing result is that most of the skill differences are nonsignificant, due to the small number of forecast cases used (40). Moreover, the skill comparisons are obscured by inherent biases due to the combination of trends, interdecadal variability and systematic errors. We use a very conservative significance testing procedure taking into account these problems. A careful examination of the skill leads to the conclusion that the statistical model performs better in the long run than the dynamical one. Of particular relevance is the question whether the valuable information contained in the empirical and the dynamical forecasts differ. If such is the case an appropriate combination of both forecasts, and/or forecasting algorithms could lead to an improvement of the skill. This is our second purpose: we propose here two hybrid procedures which combine objectively the dynamical and statistical predictive information contents. The skill of these hybrid models is compared to that of the two original models. One of the two hybrid schemes is shown to be significantly more skilful, at long lead times, than the dynamical model.