Promising subseasonal forecasting results based on machine learning

A purely statistical machine learning (ML) approach was applied to forecast near-surface temperature and precipitation anomalies over land areas in the Northern Hemisphere and Tropics. A high number of principal components (PCs) from the key variables, most importantly sea surface temperatures and the near-tropopause geopotential from reanalyses, was used as predictors to forecast the 2-weekly mean predictand anomalies in each location. Separate models were fitted for different seasons and lead times in the range of 1–6 weeks.

To select and weight the predictors and to reduce the risk of overfitting, such ML methods as least absolute shrinkage and selection operator (LASSO) regularization and ensembling based on random sampling of the predictor data were used in addition to the dimensionality reduction with PCs.

Skill analysis of the independent test sample results show that both the climatological and persistence reference forecasts were inferior compared to the ML approach on average, with all lead times, and in the majority of the target grid cells. Also, the ML approach achieved a skill that was generally comparable to the European Centre for Medium-Range Weather Forecasts (ECMWF) dynamical model.

 

Previously, these particular ML methods have been shown to work in a regional approach in Europe for seasonal time scales. According to the new results, they also work in the near-global domain and in the challenging subseasonal time scales.