Complex Systems Modelling for Statistical Forecasting of Winter North Atlantic Atmospheric Variability: a New Approach

Seasonal forecasts of winter North Atlantic atmospheric variability have until recently shown little skill. Here we present a new technique for developing both linear and non?linear statistical forecasts of the winter North Atlantic Oscillation (NAO) based on complex systems modelling, which has been widely used in a range of fields, but generally not in climate research. Our polynomial NARMAX models demonstrate considerable skill in out?of?sample forecasts and their performance is superior to that of linear models, albeit with small sample sizes. Predictors can be readily identified and this has the potential to inform the next generation of dynamical models and models allow for the incorporation of non?linearities in interactions between predictors and atmospheric variability. In general there is more skill in forecasts developed over a shorter training period from 1980 compared with an equivalent forecast using training data from 1956. This latter point may relate to decreased inherent predictability in the period 1955?1980, a wider range of available predictors since 1980 and/or reduced data quality in the earlier period and is consistent with previously identified decadal variability of the NAO. A number of predictors such as sea?level pressure over the Barents Sea, and a clear tropical signal are commonly selected by both linear and polynomial NARMAX models. Tropical signals are modulated by higher latitude boundary conditions. Both approaches can be extended to developing probabilistic forecasts and to other seasons and indices of atmospheric variability such as the East Atlantic pattern and jet stream metrics.