This post was originally published on here
MIT research scientist Judah Cohen has taken top honors in a major international weather contest, advancing a crucial piece of the climate puzzle: predicting weeks-ahead weather. His machine learning model, guided by Arctic climate signals, won first place in the 2025 AI WeatherQuest subseasonal forecasting competition run by the European Centre for Medium-Range Weather Forecasts (ECMWF).
The win spotlights a rising approach in meteorology as forecasters seek better accuracy for the two-to-six-week window. Cohen’s work could help utilities, farmers, and emergency planners prepare for swings in temperature and precipitation with more lead time.
Why Subseasonal Forecasts Matter
Weather forecasts have improved for days in advance, yet the subseasonal period often remains a blind spot. This gap affects energy demand planning, wildfire risk, crop decisions, and retail supply chains. Better skill in this range can shift how governments and businesses manage risk.
ECMWF’s competition focused on that hard middle ground. It asked participants to predict conditions weeks ahead using data science and domain expertise. The approach rewarded models that could detect signals buried in a noisy atmosphere.
Arctic Signals Meet Machine Learning
Cohen’s approach uses Arctic climate indicators, which can hint at downstream weather across North America, Europe, and Asia. These indicators can include sea ice patterns, snow cover, and shifts in atmospheric circulation that precede changes in the jet stream.
“MIT research scientist Judah Cohen is using machine learning and Arctic climate indicators to improve subseasonal weather forecasting.”
While the model’s full methods were not disclosed in detail, the strategy aligns with a growing body of research tying high-latitude variability to mid-latitude extremes. The idea is simple: if the Arctic is a driver, then its early signals can extend the usable window of prediction when paired with modern algorithms.
ECMWF’s Global Stage
ECMWF is a leading center for weather science and operational modeling. Its competitions draw teams from universities, startups, and national labs. A first-place finish suggests meaningful skill gains against baselines that are hard to beat.
“His model won first place in the 2025 AI WeatherQuest subseasonal forecasting competition, held by the European Centre for Medium-Range Weather Forecasts.”
Subseasonal contests often evaluate forecasts over multiple regions and variables. Systems must handle shifting regimes, from winter cold spells to summer heat and monsoon variability. Consistent performance across those regimes is a high bar.
Implications for Energy, Agriculture, and Public Safety
Improved subseasonal forecasts can support grid operators as they balance renewable power, heating needs, and peak demand weeks ahead. Early signals of cold snaps or heat waves can reduce costs and prevent outages.
Farmers can adjust planting, irrigation, and harvest schedules if they have more confidence in late-month conditions. Emergency managers can pre-position crews and supplies before prolonged rain or dry spells set in.
- Energy: better planning of fuel purchases and maintenance schedules.
- Agriculture: fewer losses from late frosts or sudden heat.
- Insurance: refined risk pricing and reduced catastrophe exposure.
Balancing Promise and Limits
Machine learning can detect patterns that physics-based models may miss. But it must be trained on reliable data and checked against changing climate baselines. Shifts in Arctic conditions can alter historical relationships, making continuous validation essential.
Hybrid approaches that blend statistical models with physical insight are gaining traction. They can capture known drivers, like stratosphere-troposphere coupling, while letting algorithms flag unexpected relationships.
What Comes Next
Cohen’s win signals momentum for data-driven methods grounded in climate science. The next steps include integrating these models into real-time operations, expanding regional coverage, and building user-friendly tools for decision-makers.
Further progress will depend on richer Arctic observations, open benchmarks, and careful testing during high-impact seasons. If results keep improving, the hardest weeks to predict may soon become actionable.
The latest result points in that direction, with a clear message: better signals, better models, better decisions. Stakeholders will be watching how fast these methods move from prize podiums to daily practice.
