Enhancing Seasonal Predictions with Machine Learning: A Global Perspective on SST Influence in Early Winter

Seasonal predictability of early winter (November-December) atmospheric patterns is determined, to a large extent, by the anomalous ocean surface thermal conditions. Globally, sea surface temperatures (SSTs) serve as a key driver of wintertime atmospheric patterns, with their predictive importance varying across different regions and time lags. The extratropical regions  present greater challenges for seasonal predictability due to the complexity of their atmospheric processes and interaction of signals from different sources of predictability. Globally, sea surface temperatures (SSTs) serve as a key driver of wintertime atmospheric patterns, with their predictive importance varying across different regions and time lags. In the North Atlantic region,  seasonal predictability of early winter (November-December) atmospheric patterns can be determined, to a large extent, by the anomalous ocean surface thermal conditions.  Nevertheless, current seasonal prediction systems, which rely significantly on the well known interannual phenomenon known as the El Niño-Southern Oscillation (ENSO), develop large biases in the extratropical SSTs, leading to poor performances in other key variables in those regions, such as the Euro-Atlantic region (EAR). Thus, it is important to develop alternative statistical models to overcome these problems.

This study assesses the predictive capability of global SST anomalies with lead times ranging from 1 to 10 months to forecast November-December sea level pressure (SLP) anomalies. For such purpose, we use three different statistical approaches: a Maximum Covariance Analysis (MCA) to identify dominant patterns of co-variability between SSTs and atmospheric conditions; a neural network-based method (NN) designed to capture non-linear teleconnections; and a hybrid methodology that combines the strengths of the MCA and NN techniques.

Our results highlight regions of high predictive skill across the globe, with a focus on understanding how the different initializations impact the predictability. By comparing traditional statistical methods (MCA) with advanced non-linear approaches (NN and Hybrid), this study provides a comprehensive understanding of global atmospheric predictability during early winter. In particular, significant skill in terms of anomaly correlation coefficient is found for the neural network-based methods in the EAR from 7 to 10 months in advance. Additionally, analysis of the non-stationarity of these teleconnections is found and analyzed throughout the period ranging from 1940 to 2019. Furthermore, the non-stationarity of these teleconnections over the whole period is identified and analysed, detecting windows of opportunity for more accurate seasonal forecasts. The findings aim to improve our understanding of oceanic forced atmospheric teleconnections, not only by establishing windows of opportunity for seasonal forecasts, but also by means of analysing possible drivers of these teleconnections. All of these aid in the development of more accurate and robust prediction models for managing climate-related risks worldwide.