Subseasonal-to-Seasonal Analog Forecasting Using Machine Learning and Uncertainty Quantification

Analog forecasting is an intuitive method, in which one uses past climate states that are similar to current conditions as a roadmap for prediction. However, finding these similar, analogous states is both nuanced and challenging. Given a prediction goal, e.g., temperature in the United States, are the best analogs ones that closely match the conditions in just the region of interest? The whole globe? Or some specific areas of the globe? To find the optimal analogs, we harness an artificial neural network trained on climate-model data to create a spatially-weighted map that highlights which areas of the globe are most important for selecting the best analogs. These weighted maps improve our analog forecasting accuracy and allow us to probe key modes of predictability on subseasonal-to-seasonal timescales. Moreover, by using an ensemble forecast via a group of best analogs, we utilize uncertainty quantification to extricate forecasts of opportunity.