Multi-decadal variability in predictive skill of the winter NAO

Antje Weisheimer (ECMWF and University of Oxford), Nathalie Schaller, Christopher O’Reilly, David MacLeod, James Heatley, Tim Palmer (all University of Oxford)


Results obtained from a new long dataset of ensemble seasonal hindcasts show that there is substantial multi-decadal variability in the skill of predictions of the North Atlantic Oscillation (NAO). This variability appears to be correlated with features of the large-scale circulation including the state of the NAO itself, the El Niño–Southern Oscillation (ENSO) and the Pacific Decadal Oscillation (PDO). The NAO is an irregular oscillation in atmospheric flow patterns in the North Atlantic Ocean which has a strong influence on the weather and climate in Western Europe.

Understanding the seasonal predictability of circulation anomalies over the Euro-Atlantic region has long been and continues to be a challenge. Based on skill estimates from hindcasts made over the last couple of decades, recent studies have suggested that considerable success has been achieved in forecasting the NAO during winter using current-generation dynamical forecast models. However, studies using previous-generation models had shown that forecasts of winter climate anomalies in the 1960s and 1970s were less successful than forecasts of the 1980s and 1990s. Given that the more recent decades have been dominated by the NAO in its positive phase, it is important to know whether the performance of current models would be similarly skilful when tested over periods marked by a predominantly negative NAO.

To this end, a new long dataset of ensemble seasonal hindcasts has been generated with the atmospheric part of ECMWF’s Integrated Forecasting System (IFS). This data covers the period 1900 to 2010 and is named Atmospheric Seasonal Forecasts of the 20th Century (ASF-20C). The ASF-20C hindcasts were performed with IFS Cycle 41r1, using a similar horizontal and vertical resolution as ECMWF’s currently operational seasonal forecasting System 4. The hindcasts were initialised with, and verified against, ECMWF’s atmospheric reanalysis of the 20th century (ERA-20C) and use observed sea-surface temperature (SST) and sea-ice data at the lower boundary over the ocean. The ensemble comprises 51 ensemble members. Although hindcasts were produced for all seasons, we focus here on the analysis of winter hindcasts (December–February, DJF), that is those initialised on each 1 November from 1900 to 2009. These new data provide a unique tool to explore many aspects of atmospheric seasonal climate prediction. The unprecedented length of the period covered allows for a thorough inspection of the robustness of seasonal forecast skill estimates and their variability on a timescale much longer than in previous studies.

Our findings are that the predictive skill of the NAO is non-stationary and has multi-decadal variability, with higher levels of skill during recent decades and in the early parts of the 20th century and lower levels of skill during the mid-century decades. Interestingly, the NAO index itself shows a somewhat similar behaviour, with negative values in an extended period around the middle of the century and positive amplitudes in recent decades and at the beginning of the century.

Correlation skill score and NAO index
Correlation skill score and NAO index. Time series of the correlation skill score for the ensemble mean of hindcasts of the NAO for the months of December–February initialised on 1 November, computed for overlapping 30-year windows shifted by one year at a time and plotted at the 15th year of each window; and of 30-year running mean NAO index values for the months of December–February in ERA-20C. The interannual correlation skill score for predictions of the winter NAO index over the entire period from 1900 to 2010 is significantly positive (r = 0.31, with r = 1 indicating perfect correlation and r = 0 indicating no correlation).

Interpreting the results

There is evidence for a meteorological explanation of such non-stationary predictability. We have assessed the extent to which the multi-decadal variability in skill co-varies with variations in the general circulation itself. While there is no obvious statistical reason why these two time series should be correlated, we find strong co-variations between time series of NAO skill and low-frequency time series of the NAO itself over the whole 20th century. For example, the mid-century decades of low correlation skill correspond with periods when the NAO was in a strongly negative phase. At first glance these results suggest that the model is struggling to predict the circulation in periods with negative NAO. However, it is not the case that in general the forecast model cannot skilfully predict negative NAO winters. Rather, our analysis suggests that probabilistic forecasts for strong negative and all ranges of positive NAO indices are highly skilful. Indeed, the model is remarkably good at predicting strong negative NAO winters. For example, the events associated with the highest ROC skill score are those below the 10th percentile of the climatological distribution. However, the model does not perform as well for weak negative NAO events. The overall skill in the first half of the 20th century stems from skilfully predicting a wide spectrum of NAO events.

To try to understand these results, we have studied correlations between NAO forecast skill and decadal timescale diagnostics of the general circulation further afield, such as ENSO and the PDO. The forecast skill for each of the 30-year periods does seem to be related to the dominant phase of El Niño, with periods of positive SST anomalies in the central tropical Pacific coinciding with periods of strong NAO skill in the hindcasts. This is also the case for the PDO, which is in a positive phase during periods when the hindcasts are most skilful. These correlations with the general circulation suggest that the decadal variations in NAO forecast skill are indeed linked to the general circulation. More specifically, the NAO forecast skill seems to be correlated most strongly with variations in SSTs in the tropical Western Pacific and the Indian Ocean. Further numerical experiments are needed to understand such teleconnections.

As explained in greater detail in an article by the authors in the Quarterly Journal of the Royal Meteorological Society (doi:10.1002/qj.2976), during the mid-century period of low skill the DJF NAO exhibits remarkable persistence from the November NAO. It is further found that the decades of high NAO persistence from the 1940s to the 1970s coincide with periods of enhanced intra-seasonal variability of geopotential height at 500 hPa over the Atlantic sector. These findings are consistent with the hypothesis that upper-level Rossby wave-breaking events occur more frequently during periods of negative NAO than during periods of positive NAO, which will be a focus of future research.

In conclusion, the mid-century decades stand out as an important period on which to test the performance of future seasonal forecast systems. Achieving good forecast skill for the more recent decades with predominantly positive NAO indices is not sufficient to guarantee similarly good performance for decadal periods with negative NAO index, which might occur again in the future. Our findings underline the importance of a representative re-forecast dataset for robust conclusions about the levels of predictive skill in predicting the Atlantic-European climate in the future.