Fifty people look at the ocean. Each notices something different: one sees warm water, another wind, a third depth. Individually they make mistakes, but if you combine their opinions and trust those who have more often been right, you get a council of sages. That is how a new method for forecasting El Niño works—the phenomenon in which the Pacific Ocean overheats and changes weather across the entire Earth.

Instead of people, the council consists of fifty simple mathematical models. Each is trained on satellite images of the ocean since 1981 and looks at its own set of features: surface temperature, wind, heat at depth. To prevent models from overestimating themselves, the scientists used entropy training—a way to force the algorithm to distribute probabilities as cautiously as possible until data prove otherwise.

The main secret is the elder of the council, a meta-model. It monitors each of the fifty models: how often it erred, in which season it performed better, how confident its answers were. The elder then weighs the votes: if a model is weak in June, its June forecast is almost disregarded. In a validation from 2012 to 2022, such a council predicted the El Niño phase (El Niño, La Niña, or neutral state) 24 months ahead—twice as far as typical operational forecasts. And the energy used for the calculation was equivalent to two hours of laptop operation, whereas supercomputer models consume thousands of times more.