In This Article
Microsoft has introduced an open-source machine-learning tool designed to make weather forecasts more reliable two to six weeks in advance. Rather than building a new forecasting system, it corrects systematic errors in existing models. In a research study, the approach improved results from leading European forecasting systems while requiring only standard CPU computing resources.
Key takeaways
The tool is designed to complement existing forecasts, not replace them, and its reported gains come from testing against specific systems and targets.
- Probabilistic Bias Correction (PBC) identifies and adjusts recurring forecast biases.
- In the study, it improved precipitation accuracy for 98 percent of evaluated targets and temperature accuracy for 92 percent.
- It also doubled the accuracy of the European Centre for Medium-Range Weather Forecasts’ AI system in the researchers’ analysis.
- The code is open source and can run on CPUs, potentially lowering barriers for agencies with limited computing resources.
How the correction layer works
Forecast systems—whether based on atmospheric physics or machine learning—often become less dependable beyond roughly two weeks. Models use approximations to address gaps in observations and scientific understanding; errors from those assumptions can accumulate as forecasts extend further into the future.
PBC applies machine learning to identify systematic bias in an existing model’s output and adjust the resulting probabilities. For example, a model relying heavily on historical observations may not fully reflect a warming climate. Correcting that bias can improve its output without discarding the underlying forecast signal.
Promising results, with operational work ahead
The study’s results point to a potentially useful way to extend forecast reliability, but performance in a research evaluation does not guarantee the same gains in day-to-day operations or in every region. ECMWF’s machine-learning lead described the research as valuable to the wider community while noting that putting a method into an operational forecasting process remains challenging.
PBC also needs an existing forecast model to work on. Local agencies would need to configure and test it against regional data before relying on it. A project involving nonprofit Rhiza Research and Kenya’s meteorological service is examining PBC among other methods for improving early warnings.
What data teams can take from the research
The broader lesson for enterprise leaders is that better decisions do not always require replacing an entire system: targeted correction and reliable data integration can improve the value of existing tools. Organizations that bring external forecasts into operational dashboards need clear pipelines, consistent measures and appropriate validation. Spargent helps enterprises build and scale data platforms with Microsoft Fabric and Power BI, supporting the data foundations needed to turn complex inputs into timely, decision-ready reporting.
