The forced response and decadal predictability of the North Atlantic Oscillation: non-stationary and fragile skills

Abstract We investigate the forced response of the North Atlantic Oscillation (NAO) – calculated as the ensemble mean – in different large ensembles of climate models includingsimulations with historical forcings and initialized decadal hindcasts. The forced NAO in the CMIP6 historical ensemble correlates significantly with observations after 1970. However, the forced NAO shows an apparent non-stationarity with significant correlations to observations only in the period after 1970 and in the period before 1890. We demonstrate that such apparent non-stationarity can be due to chance even when models and observations are independent. For the period after 1970 the correlation to the observed NAO continues to increase while the amplitude of the forced signal continues to decrease – although both with some signs of saturation – when the ensemble size grows. This behavior can be explained by a simple model assuming a very small signal-to-noise ratio in the models. We find only rather weak evidence that initialization improves the skill of the NAO on decadal time-scales. The NAO in the historical ensembles including only natural forcings, well-mixed greenhouse-gases, or anthropogenic aerosols show skills that are not significantly different from zero. The same holds for a large single-model ensemble. The skills of these ensembles, except for the well-mixed greenhouse-gas ensemble, are also significantly different from the skill of the larger full historical ensemble even though their ensemble sizes are smaller. Taken together, our results challenge the possibility of useful NAO predictions on decadal time-scales.