Extreme multi-decadal trends in the North Atlantic Oscillation

Abstract

Stochastic processes are shown to be useful tools for quantifying extreme trends in climate indices. The variance of the trend distribution is shown to generally increase with autocorrelation, with an increase in extreme trend exceedance probabilities. The winter North Atlantic Oscillation (NAO) index has weak autocorrelation which is underestimated in historical climate models and helps to explain the underestimation of extreme trends. The maximum observed 31-year NAO trend occurred in 1963-1993 and is estimated to have a 1 in 20 chance of being exceeded in the 144-year historical record using fitted stochastic models. Climate models and stochastic models without autocorrelation underestimate this probability as a 1 in 200 chance. The NAO trend in the 1963-1993 window was identified due to its unusual nature. If this window is wrongly treated as a randomly chosen single window, the exceedance probability is further reduced (a 1 in 1000 chance). Post-processing methods are proposed to increase the low autocorrelation in climate models and are shown to improve the simulation of extreme trends and also increase the variance of ensemble mean trends. Future projections show a small systematic increase in end-of-century NAO ensemble mean trends relative to the magnitude of the radiative forcing. The probability of a maximum 31-year trend greater than that observed is 3 7% in the next 75-years (under the higher “business as usual” radiative forcing scenario), which is similar to the historical model probability for the last 75-years. Near-term projections of the next 31 years (2024-2054) are relatively insensitive to the scenario, showing no forced trend in the models but a large ensemble range due to internal variability ( 7.41 to 7.68 hPa/decade) which could increase or decrease regional climate change signals in the Northern Hemisphere by magnitudes that are underestimated when using raw climate model output.