Overview
- The research team published their results in Nature on Wednesday and released the tool as open-source software that can be run on ordinary CPUs.
- Their Probabilistic Bias Correction (PBC) is a postprocessing ‘wrapper’ that learns systematic errors in existing models and adjusts forecasts rather than replacing the models themselves.
- Applied to ECMWF’s leading systems, PBC roughly doubled forecast accuracy in the two-to-six-week range and outperformed other entries in ECMWF’s 2025 real-time competition.
- The method substantially improved forecasts for extreme events and floods, with reported flood‑prediction skill gains in many tests ranging from roughly 76% to 150% for the most severe cases.
- The work challenges the long‑held ‘two‑week’ predictability limit by making the subseasonal range usable for farmers, energy planners, and disaster managers and is already being tested by resource‑limited services such as Kenya’s meteorological office.