In the hall, fifty sages confer, each observing the ocean in their own way. One listens to surface temperature, another to the wind, a third to how heat hides in the depths. Their judgments often contradict each other, but there is an elder who knows which of them to trust more. That is how the new El Niño forecasting method works, for the climate cycle that reshapes weather across the planet every few years.

The sages in this problem are not people but fifty mathematical models trained on satellite observations since 1981. Each model views the ocean through principal components, compressed descriptions of how temperature, wind, and the deep layer change. Rather than relying on a single model, the authors assembled an ensemble and applied entropy-based training, an algorithm that avoids overconfidence by keeping probabilities as uniform as possible until the data force it to sharpen.

The key trick is the meta-model, the elder in the council. It is trained on thirty characteristics of each base model: how often the model erred, how confident it was, in which season it performed best. Based on these features, the meta-model estimates the probability that each forecast is correct and weighs the votes. If a sage is weak in June, their voice is barely heard in June.

The results are impressive. In retrospective tests from 2012 to 2022, the ensemble retained skill up to 22–24 months ahead, twice as far as operational forecasts from the International Research Institute for Climate and Society (IRI). The model predicted the strong El Niño events of 2015/16 and 2018/19, as well as several La Niña episodes, with a 24-month lead. Meanwhile, a full 24-month forecast computation consumes about 0.113 kWh, comparable to running a laptop for two hours, whereas a single run of a dynamical climate model is akin to the monthly electricity use of an average apartment.

The council of sages does not replace expensive dynamical models but complements them. The low computational cost makes the method accessible to weather services in developing countries that lack supercomputers. And the ability to learn solely from real observations, without inheriting simulation errors, paves the way for forecasts that honestly say, "I am not sure." In a world where climate is becoming less predictable, such honesty is worth more than overconfidence.