An artificial intelligence method improves weather forecasts up to six weeks in advance

Predicting the weather on a sub-seasonal scale (two to four weeks) using traditional physical models is currently not very feasible because the inherent uncertainty in observations rapidly degrades forecasts as the days go by. The study led by Hannah Guan, published in Nature, presents a new machine-learning-based data post-processing tool (Probabilistic Bias Correction, or PBC) that significantly improves sub-seasonal forecasts.

On a global scale, the study doubles the performance of the artificial intelligence (AI)-based forecasting system and improves it by more than 90% in traditional physical models. However, it should be noted that the positive impact is concentrated primarily in the tropical belt and remains quite limited in Europe, for example. The study highlights that the PBC method can also help improve the prediction of extreme events more than two weeks in advance. This is the primary goal of sub-seasonal forecasting: to anticipate events such as frosts, droughts, or storms weeks in advance, giving the agricultural, energy, and emergency sectors the time they need to take action.

One of the advantages of the PBC method is that it is not limited to correcting an average value; rather, it adjusts both the most probable result and the uncertainty associated with it. To do so, it learns from past errors and takes into account recent weather trends and the typical climate of each location. It is a lightweight method that can be updated with new data without having to rerun the entire weather model, although, as we have already mentioned, it always needs to rely on a model and does not function independently. Its value also lies in the fact that it has been tested not only with historical data but also with real-time forecasts, before the actual weather conditions were known. Furthermore, improvements are evident in different regions and times of year, even in models that had already been corrected previously.  
  
The results of this study are promising, although the method remains in the experimental phase and has been tested primarily using historical weather reconstructions. Therefore, it is advisable to view the large percentage improvements with caution and to continue studying how it performs in other regions, at finer scales, and in the face of extreme weather events.

The significance of this work lies in the use of probabilistic post-processing (PBC), which learns to correct the systematic errors in the best existing models—both dynamic and AI-based—without modifying them. It is noteworthy that the PBC method won the ECMWF’s international AI Weather Quest competition in 2025 for all variables and time horizons, outperforming the operational models of six national meteorological centers (ECMWF, NOAA, JMA, KMA, ECCC, HMCR). The most significant improvement is seen in precipitation for the 3- and 4-week time horizons, where the models’ direct output exhibits a very high systematic bias relative to the climatological reference. The code is open-source, and the approach is applicable to any weather prediction model, which greatly expands its practical scope. 

It should be noted, however, that a new atmospheric model is not built from scratch. The PBC method is a post-processing tool that must be applied to the base models. If the dynamic models improve, PBC can adapt, but if they fail in certain complex and rare situations, the PBC method cannot correct this. Another limitation of the study is that the validation uses ERA5 as the reference truth, which has its own uncertainties since it is another model that simulates precipitation (it does not assimilate actual precipitation observations, but it does assimilate temperature, humidity, and wind).  

On the other hand, although the leap in precipitation predictive ability is notable (it goes from negative to positive ability), the relative improvement is misleading because it is akin to “dividing by zero.” In other words, “doubling the ability” from a modest starting point does not necessarily mean it is a “useful forecast.” The subs seasonal predictability window remains inherently challenging, ensuring many more years of work in mathematical and meteorological research.

The ECMWF AI Weather Quest [a competition won by a model that incorporated PBC] sets out to stimulate new ideas and encourage teams around the world to exchange results and learn from one another in an open, collaborative way. This paper [Advancing Subseasonal Forecasting with Machine Learning] is a good example of new ideas being developed for the benefit and consideration of the community. As always, turning ideas into reality an operational process is a most challenging step.