Understanding the Predictability of the Winter North Atlantic Oscillation using Dynamical Seasonal Forecast Models and Machine Learning Techniques

Oral Presentation 

Current generation dynamical seasonal forecast models have been shown to be capable of forecasting the winter North Atlantic Oscillation (NAO) with significant skill. However, there is some uncertainty in the skill of these forecasts. This is seen both by intermittency of skill in individual years within hindcast periods, and also in the decadal variability of skill over long hindcasts. 

The Nonlinear AutoRegressive Moving Average with eXogenous inputs (NARMAX) systems identification approach is an interpretable machine learning method which can be used to identify and model linear and non-linear dynamic relationships meteorological and related variables, including identifying non-stationary associations that can occur over time. NARMAX has been shown to produce skilful forecasts of the NAO when trained on historical observation data. A key element of NARMAX is that information about the predictors used in its predictions and their relative importance is retained, and can therefore be used as a tool to understand more about the key processes driving NAO variability. 

Here we use NARMAX in conjunction with dynamical seasonal forecast models to understand more about the predictability of the NAO and the temporal intermittency of this predictability. We analyse the key predictors identified by NARMAX in periods where the NAO is well forecast and poorly forecast by the dynamical seasonal forecast models. In addition, we apply NARMAX to the dynamical model NAO hindcasts. We compare the predictors identified by NARMAX in this case with those identified when NARMAX is applied to observations. These results will help to understand how NAO predictability differs between the “real world” and the “model world”, and identify potential deficiencies in the dynamical seasonal forecast models.