Extreme temperatures in the U.S. cause over $100 billion in damages each year, making it crucial to improve predictions of these temperature extremes for the upcoming weeks (2-4 weeks ahead). This study examines three machine learning methods, including Random Forests (RF) and Echo State Networks (ESN), to forecast weekly average temperatures in five regions of the U.S. The machine learning methods are also compared against traditional forecasting methods.

The results show that Random Forests perform the best overall, especially in predicting extreme temperatures. Echo State Networks perform well for short-term forecasts but not as effectively for longer ones. The study also analyzes which factors influence the models' predictions, finding that different variables are important depending on the region and how far ahead the forecast is. Overall, the findings highlight the complexity of predicting temperatures and the effectiveness of machine learning methods in this context.