In 1997, the Pacific Ocean off the coast of Peru warmed so much that fish disappeared and torrential rains hit the desert. This is El Niño, a warm current that every few years changes weather across the planet. Predicting its arrival six months ahead is already possible, but looking further is like guessing a ship's route from a single point on a map.

However, the climate system has its own map: the attractor. This is the shape traced by the system's trajectory in state space when the axes are not time but the variables themselves: pressure, temperature, wind speed. For El Niño, that map is the Southern Oscillation Index (SOI), the pressure difference between Tahiti and Darwin, recorded monthly since 1866. Each El Niño event leaves its own loop on this map, much like a ship's wake.

The scientists decomposed 135 years of records into nine layers, from fast intraseasonal oscillations to a slow trend, and built a nine-dimensional space. Then, for the 1997–1998 event, they took its starting point and searched the past for similar points. Three were found: 1911, 1972–1973, and 1982–1983. Shifting their trajectories to the event's start and averaging them yields a forecast. The error two years ahead was less than half the standard deviation of the entire signal, and four years ahead no more than 0.37. This means that the map of the past could have revealed the outlines of the future in advance.

But the map does not always work. For the 1982–1983 event, no suitable loop was found in the past, leaving the forecast inaccurate. Perhaps this is a rare state of the attractor that has not repeated in 135 years of observations. Nevertheless, for most events of the 20th century, the method gives an average forecast two to four years ahead, and this is shown for the first time for the extreme El Niño of 1997–1998, which was usually poorly predicted.

Previously, it was thought that looking beyond six months was nearly impossible because of too much noise in the data. Now it is clear that the SOI index itself encodes nonlinear information sufficient for forecasts years ahead. The method requires no complex models, only wavelet decomposition and a search for similar orbits. It can be applied to other climate records to test whether predictability persists beyond the 20th century. The map of the past, if you learn to read it, shows the road to the future.