Probabilistic Bias Correction enhances two-to-six-week forecasting, offering significant advantages for farmers, resource managers, and disaster responders.
Key Highlights
- Current weather forecasts demonstrate persistent unreliability within the two-to-six-week timeframe, creating a notable void for decision-makers across diverse sectors and governmental entities.
- Probabilistic Bias Correction (PBC) emerges as an innovative machine-learning framework conceived by Microsoft Research, designed to rectify historical inaccuracies in weather models to enhance forecast precision.
- In preliminary assessments, PBC excelled in a global contest aimed at refining extended-range (sub-seasonal) forecasting. Its capabilities are notably beneficial in practical scenarios, such as enhancing crucial monsoon predictions for Kenyan farmers.
Modern weather forecasting has achieved remarkable reliability in two-week outlooks, and long-term climate forecasts are progressively accurate.
Nevertheless, a significant chasm remains in sub-seasonal forecasting between two and six weeks, where existing models fall short.
Improved predictions within this crucial window would facilitate essential decisions related to agriculture, wildfire management, water and energy allocation, and preparations for extreme weather events.
Figure 1: On January 29, 2026, MicroDuet, utilizing PBC, was able to predict the catastrophic North American Blizzard of 2026 (February 22-24), one of the most severe winter storms recorded in specific U.S. locales.
For years, the collaborative efforts of Microsoft researchers and external partners have been directed at resolving this quandary.
Rather than developing yet another competing model in a saturated market, the focus shifted towards addressing known errors in existing models.
This novel strategy, labeled Probabilistic Bias Correction (PBC), has shown promising potential in delivering more actionable— or skillful— forecasts for temperature, precipitation, and atmospheric pressure.
Traditional physics-based models assess the current weather and simulate future atmospheric conditions using scientific principles governing the behavior of the atmosphere and oceans.
In contrast, contemporary AI models leverage vast data for training. Both methodologies perform admirably for short-term forecasting; however, neither achieves reliable results within the two-to-six-week period due to accumulated errors and the intricate interactions between global and local climatic systems.
“Our innovative pipeline (PBC) accepts any preferred model as input, learns its error patterns, and subsequently rectifies those shortcomings, resulting in superior output forecasts for the sub-seasonal timeframe.”
– Lester Mackey, Senior Principal Researcher at Microsoft Research New England
A recent study (opens in new tab) published in the scientific journal Nature elaborates on how PBC employs a streamlined machine-learning system to identify systematic biases in any model’s historical forecasts.
Remarkably, PBC has enhanced the performance of the world’s foremost operationally debiased sub-seasonal model from the European Centre for Medium-Range Weather Forecasts (opens in new tab) (ECMWF) across 98% of precipitation, 92% of temperature, and 91% of atmospheric pressure targets, effectively doubling the accuracy of ECMWF’s AI Forecasting System (AIFS).
In August 2025, ECMWF initiated the AI Weather Quest (opens in new tab), a real-time competition aimed at enhancing probabilistic sub-seasonal forecasting.
The MicroEnsemble team, comprising researchers from Microsoft and external collaborators, submitted a PBC-based solution, MicroDuet, which applied PBC to ECMWF’s leading dynamical model and a hybrid AI-dynamical model, PoET (opens in new tab), designed to enhance the probabilistic skills of forecast ensembles via transformer techniques.
Throughout the inaugural three-month season, MicroDuet secured first place for all weather variables and lead times, surpassing six leading operational forecasting centers and 34 international forecasting teams.
Furthermore, MicroDuet demonstrated the highest accuracy across an entire year of real-time forecasting, notably predicting two major winter storms (opens in new tab) affecting the United States nearly three weeks in advance.
Figure 2: On November 20, 2025, MicroDuet, powered by PBC, predicted the severe cold air outbreak in the Eastern United States, an extreme event overlooked by the leading debiased dynamical model from ECMWF.
The AI Weather Quest serves as an exceptional real-time benchmarking platform for sub-seasonal forecasts.
The contributions from Team MicroEnsemble underscore how machine learning can augment the precision of traditional dynamical forecasting systems through post-processing, stated Dr. Joshua Talib, a scientist specializing in sub-seasonal forecasting at ECMWF.
In tandem, ECMWF is advancing predictive systems and tools aimed at enhancing the skill, reliability, and usability of sub-seasonal predictions, while also progressing in wholly data-driven forecasting methodologies.
Figure 3: On January 1, 2026, MicroDuet, leveraging PBC, anticipated Winter Storm Fern (shown in satellite imagery), a severe ice and snow storm occurring from January 23-27, reported to cause 174 fatalities and accumulate $4 billion in damages.
Real-world Implementation of PBC
Beyond the confines of the laboratory, PBC is addressing substantial real-world challenges, particularly in East Africa.
Genevieve Flaspohler leads Rhiza Research (opens in new tab), a nonprofit organization dedicated to enhancing AI-driven weather forecasting.
She collaborates with the Kenya Meteorological Service Authority (KMSA) (opens in new tab), collaborating with experts from KMSA, the Climate Hazards Center (CHC) at the University of California, Santa Barbara, and Germany’s Karlsruhe Institute of Technology (KIT) to produce forecasts crucial for the success of smallholder farmers.
Figure 4: Climate Hazards Center field scientist Johnson Barasa gathers data in Kenya following below-average long rains in 2026 that adversely affected maize yields.
“A precise weather forecast, conveyed at the opportune moment through appropriate communication channels, can significantly influence agricultural outcomes,” remarked Flaspohler.
“This is particularly pertinent in East Africa, which has experienced dramatic fluctuations between record dry spells and unprecedented wet seasons.”
Flaspohler is aiding Kenyan meteorologists in evaluating and implementing AI modeling techniques, including PBC, as one of Kenya’s rainy seasons commences.
Accurate forecasts are pivotal in a region largely dependent on rainfall, lacking irrigation infrastructure.
Decisions regarding seed purchases, planting timings, and fertilization schedules hinge upon climatic predictions, with the ideal scenario being a period of consistent, gentle rain devoid of severe drought or torrential downpours that could devastate young crops.
“With the current El Niño phenomenon being the warmest on record, significant apprehension exists among governments and the climate science community concerning its potential implications,” Flaspohler elaborated.
“An urgent demand for robust and straightforward forecasting methods is apparent.” Figure 5: An example of rainfall and temperature forecasts issued by the Kenya Meteorological Service Authority. Image credit: Kenya Meteorological Service Authority
During this rainfall season, PBC is assisting the KMSA team in identifying potential instances of extreme rainfall as they prepare weekly forecasts for the country’s meteorological service.
Dr. Phillip Okello, a principal meteorologist at KMSA who has assessed PBC, is optimistic about its capabilities.
“Traditional bias correction methods merely modify the forecasted quantity of rainfall. PBC, however, rectifies the underlying probabilities, achieving greater forecast accuracy across the Greater Horn of Africa than any other approach we evaluated.
The subsequent phase involves refining its predictions of heavy rainfall and droughts, which are critical for agricultural stakeholders.”
— Phillip Okello, Principal Meteorologist, Kenya Meteorological Service Authority
Amid the unpredictable weather patterns of recent years, Flaspohler emphasizes the urgency of this initiative.
“As precipitation becomes increasingly difficult to forecast, we must constantly enhance our methodologies,” she remarked.
“This necessitates the continuous rollout of novel techniques and superior scientific insights to furnish farmers with the most effective information available.”
The burgeoning success of PBC indicates that substantial advancements in forecasting accuracy may be feasible. Although still early in the agricultural season, initial feedback appears promising.
Anticipation for rainfall is palpable among the Kenyan populace, and PBC’s specialized sub-seasonal forecasts, tailored to enhance local performance, will enable regional forecasters to optimize global modeling outputs for their specific country.
“This has significantly improved our preparedness for the current rainy season and facilitated information dissemination to farmers,” Flaspohler concluded.
Figure 6: PBC is aiding the team in assessing the likelihood of substantial rainfall as they prepare weekly updates for Kenya’s meteorological service.
Future Directions for PBC
Given the promising outcomes associated with PBC in both experimental assessments and initial field implementations, Microsoft is actively investigating further applications and enhancements of this technology.
This includes identifying sectors that could greatly benefit from PBC-related tools, as well as determining which weather models could gain substantial improvements from its machine-learning corrections.
These explorations are likely to generate new research inquiries, fostering additional advancements in the technology.
One notable instance is Microsoft Weather’s sub-seasonal-to-seasonal (S2S) model, which offers weekly probabilistic guidance extending up to six weeks.
This model has transitioned into real-time operations and is currently in private preview, accessible to select clients for testing as an integral part of Microsoft Weather’s comprehensive suite of meteorological solutions.
The team behind Microsoft Weather, employing AI-driven forecasting strategies, supplies weather functionalities across platforms such as Windows, Bing, Copilot, Edge, MSN, and other Microsoft offerings, delivering high-quality meteorological data instrumental for enterprises in making pivotal operational and business-related decisions.
Interested parties looking to explore the S2S model for commercial or organizational usage are encouraged to reach out to the Microsoft Weather team.
Moreover, PBC is currently available as open-source software (opens in new tab), affording researchers and developers the opportunity to explore, test, and expand upon the framework.
The aim is to stimulate continued collaboration with the broader forecasting community to further propel both the technology and its potential applications.
Figure 7: Members of the MicroEnsemble team showcased PBC at the AI Weather Quest Annual Awards Webinar, achieving top accolades at the global competition organized by the European Centre for Medium-range Weather Forecasting.
Source link: Microsoft.com.
