Applying Machine Learning to Improve Subseasonal-to-Seasonal (S2S) Forecasts

The discipline of seasonal climate prediction began as an exercise in simple statistical techniques. However, today the large government forecast centers almost exclusively rely on complex fully coupled dynamical forecast systems for their subseasonal to seasonal (S2S) predictions while statistical techniques are mostly neglected and those techniques still in use have not been updated in decades. We argue that new statistical techniques mostly developed outside the field of climate science, collectively referred to as machine learning, can be adopted by climate forecasters to increase the accuracy of S2S predictions. As proof-of-concept our team participated in the Subseasonal Climate Forecast Rodeo managed by the U.S. Bureau of Reclamation and NOAA. The Subseasonal Rodeo was a year-long real-time forecasting challenge, in which participants aimed to skillfully predict temperature and precipitation in the western U.S. two to four weeks and four to six weeks in advance. To meet this challenge, we used both observational data and dynamical model forecasts to develop a machine learning-based forecasting system and a publicly-available Subseasonal Rodeo dataset suitable for training and benchmarking subseasonal forecasts. Our system significantly outperformed the contest baselines—a debiased 32-member CFSv2 forecast and a damped persistence statistical forecast—with skill exceeding that of the top contest participant for each target variable and forecast horizon. We are updating the model and expanding it to cover the entire U.S. and plan to present new results.