Statistical Learning Methods as a Basis for Skillful Seasonal Temperature Forecasts in Europe

Abstract

Abstract A statistical learning approach to produce seasonal temperature forecasts in western Europe and Scandinavia was implemented and tested. The leading principal components (PCs) of sea surface temperature (SST) and the geopotential at the 150-hPa level (GPT) were derived from reanalysis datasets and used at different lags (from one to five seasons) as predictors. Random sampling of both the fitting years and the potential predictors together with the Least Absolute Shrinkage and Selection Operator regression (LASSO) was used to create a large ensemble of statistical models. Applying the models to independent test years shows that the ensemble performs well over the target areas and that the ensemble mean is more accurate than the best individual ensemble member on average. Skillful results were especially found for summer and fall, with the anomaly correlation coefficient values ranging between 0.41 and 0.68 for these seasons. The correct simulation of decadal trends, using sufficiently long time series for fitting (70 years), and the use of lagged predictors increased the prediction skill. The decadal-scale variability of SST, most importantly the Atlantic multidecadal oscillation (AMO), and different PCs of GPT are the most important individual predictors among all predictors. Both SST and GPT bring equally much predictive power, although their importance is different in different seasons.