A low-cost post-processing technique improves weather forecasts around the world

Current computer-generated weather forecasts utilize gridboxes of approximately 400 km², predicting a single value per box. This approach often fails when weather conditions vary significantly within a gridbox. This study introduces a statistical post-processing method for ensemble forecasts that accounts for sub-grid variability, gridbox-scale bias, and weather dependence. Global application of this method, termed ecPoint, demonstrates substantial skill improvements, particularly in forecasting extreme rainfall, extending useful forecasts to 5 days ahead compared to less than 1 day without post-processing. The improvement is attributed to large calibration datasets created by globally aggregating forecast-observation differences for similar weather types and a strong focus on meteorological understanding. Potential applications include improved flash flood warnings, insights into model weaknesses, and global pointwise re-analyses.

Timothy David Hewson, Fatima Maria Pillosu