Predictability of Common Atmospheric Teleconnection Indices Using Machine Learning

Abstract Medium-range weather forecasting continues to be an elusive challenge in the field of meteorology. The ability of dynamic weather models to represent changes in atmospheric variability at time scales on the order of months is limited, such that meteorologists typically resort to climatic scales to assess predictability. Atmospheric teleconnections provide an opportunity to assess interconnectedness at a planetary scale and have known relationships with seasonal features, particularly in wintertime temperatures and precipitation. However, these teleconnections are difficult to predict owing to their chaotic nature. In this study, the predictability of the cold-season monthly phases of four common teleconnection indices, the North Atlantic Oscillation (NAO), the Pacific North American oscillation (PNA), the West Pacific Oscillation (WPO), and the Arctic Oscillation (AO) is quantified using multiple machine learning methods. Lagged indices out to 6 months are employed for all four teleconnection patterns are used to predict each upcoming teleconnection index. Improvements offered by support vector machines and random forests are compared against a baseline multivariate linear regression to quantify the benefits of applying machine learning to the research problem. Ultimately, the resulting models offer a first attempt at quantifying monthly atmospheric variability using machine learning and are an important first step towards improved medium and long-range weather forecasting.