Week 3-4 Predictions using Machine Learning Techniques

Predictions beyond the 2-week limit of atmospheric predictability would be extremely valuable to society if they could be done reliably. For instance, such predictions could provide early warning to heat waves and cold spells as well as droughts and excessive precipitation. In this talk, I will describe statistical models found through machine learning techniques that can skillfully predict the average temperature and precipitation over the US 3 to 4 weeks in advance. The predictors of these models are sea surface temperatures of the Atlantic and Pacific ocean. A novel part of this work is that the statistical models were trained using long climate simulations. In particular, I will show that such models can produce skillful predictions in other climate model simulations as well as in observations. The features of the sea surface temperatures that make these predictions skillful will be discussed. This work shows that, in principle, climate models can produce operational forecasts in the week 3-4 range.