The ocean remembers its weather. Each month, its surface temperature forms a vast map, and if you find a similar map in the past, you can see what happened next. That is exactly how the new method for forecasting El Niño works: the warm current that, every few years, alters the climate of half the planet.

Previously, forecasts relied on linear models. They work well when everything changes smoothly and proportionally: slightly warmer water, slightly stronger winds. But El Niño does not behave that way. Its birth involves nonlinear interactions that a linear model cannot see in principle. Hence, its accuracy drops after six months, and beyond that the forecast becomes useless.

The new method belongs to the nonparametric family: it does not try to guess the formula by which the ocean changes. Instead, it uses the observational history from 1870 to 1997, then tests itself on the following twenty years. The result: a useful forecast (correlation above 0.5) persists for 9–14 months depending on the decade, whereas a linear model loses skill by 6–7 months. On data from the CCSM4 climate model, the advantage is even more pronounced: 24 months versus 18.

Especially important is the "spring predictability barrier": the traditional drop in forecast accuracy when crossing spring. The new method softens this barrier: in June–September its correlation stays around 0.5–0.6, while the linear model's falls to 0–0.3. For the historical El Niño of 2015/16, a forecast initialized in June 2015 predicted a strong event peaking in December, underestimating the amplitude by only 0.4°C.

Forecasting from the ocean's memory is not an attempt to replace expensive dynamical models, but a way to extract the maximum from already accumulated data. The method learns from a relatively short observational record and provides a theoretically optimal forecast as a conditional expectation. It can be extended to other geophysical phenomena with nonlinear dynamics, such as forecasting monsoons or North Atlantic variability. The ocean remembers, and now we are better at reading its memory.