Study Published in *Nature* Improves Forecasting of Extreme Events
Forecasting the weather several weeks in advance is one of the greatest challenges in meteorology and is essential for anticipating extreme events. An article published this Wednesday (the 14th) in the journal *Nature* presents a set of statistical correction techniques that nearly double the accuracy of forecasts over the so-called subseasonal horizon—two to six weeks. Paulo Orenstein, a researcher at IMPA (Institute of Pure and Applied Mathematics), is the only Brazilian among the authors, who are affiliated with institutions such as Harvard, MIT, and the University of Toronto.
“Predicting the weather far in advance is difficult because small errors in atmospheric modeling multiply as days and weeks go by. We have developed a set of statistical correction techniques that prevent these errors from accumulating. The correction can be applied to any existing climate model, and because it is probabilistic, it improves not only the forecast itself but also our ability to predict the probability of extreme events,” explains Orenstein, a co-author of the study.
This finding applies both to classical climate models—based on atmospheric simulations and differential equations—and to the most recent models based on artificial intelligence (AI). No matter how good the model may be, minute uncertainties in the initial state of the atmosphere grow over time and, combined with systematic errors in each forecasting system, create a “predictability desert” in the intermediate weeks—the gap between the range of short-term models and the seasonal scale.
This is where the method presented in the article—called PBC (Probabilistic Bias Correction)—comes in. PBC uses historical forecasts to learn where and how each model typically makes errors and systematically corrects these deviations. The correction combines two complementary strategies: Debias++, which adjusts for errors specific to each location, date, and intensity range, and Persistence++, which incorporates recent observations and climatology throughout the forecast.
Since the correction is probabilistic in nature, the benefit goes beyond point estimates: the method also improves the ability to estimate the probability of more or less extreme events occurring—information that is essential for anticipating droughts, torrential rains, heat waves, or cold snaps.
The researchers evaluated the method using three meteorological variables: temperature, precipitation, and atmospheric pressure on the subseasonal timescale. The PBC doubled the predictive ability of the ECMWF’s (European Centre for Medium-Range Weather Forecasts) AI-based forecasting system and improved the performance of dynamically modeled systems—which had already been operationally corrected—for 91% of pressure targets, 92% of the temperature targets, and 98% of the precipitation targets. According to the authors, essentially all climate models currently in use benefit from these techniques.
Researchers Win Global Climate Prediction Challenge
Those same ideas were behind the victory of the team that developed the MicroDuet model at the AI Weather Quest, an international competition organized by the ECMWF over the course of 12 months. The model took first place in all variables and timeframes evaluated, outperforming 34 teams from around the world, the operational models of six government forecasting centers, and the AI systems developed by the ECMWF itself. The European center plans to integrate the award-winning models and test them in real time.
Among the practical applications of this breakthrough is the ability to forecast extreme events weeks in advance—valuable information for agricultural planning, water and energy management, and civil defense. The next step, according to the researchers, is to apply the models to real-world variables, with a direct impact on agricultural and planting decisions.
The article “Advancing Subseasonal Forecasting with Machine Learning” was authored by Hannah Guan, Soukayna Mouatadid, Paulo Orenstein, Judah Cohen, Haiyu Dong, Zekun Ni, Jeremy Berman, Genevieve Flaspohler, Alex Lu, Jakob Schloer, Joshua Talib, Jonathan A. Weyn, and Lester Mackey, with affiliations at Microsoft Research, Microsoft, Harvard, the University of Toronto, IMPA, MIT, Atmospheric and Environmental Research, Rhiza Research, and ECMWF.