The Pacific Ocean breathes, and that breath has a pulse. Every few years, its surface warms more than usual, bringing torrential rains to Peru and drought to Australia. This is El Niño, the warm phase of the ocean temperature oscillation. Predicting it in advance has been attempted for decades, but most models stumble over the "spring barrier": in spring, the ocean and atmosphere are so unstable that forecasts lose their skill.
One way to bypass the barrier is to listen not to individual points but to the whole chorus at once. A climate network is built as follows: 14 points in the El Niño basin and 193 points outside it are connected if their temperatures changed synchronously over the past 200 days. The strength of each link is the signal-to-noise ratio: how coherently two points warmed and cooled, compared with their usual chatter. If the average strength of all links crosses the threshold of 2.82, an alarm goes off one year before the event.
The test ran in real time for 12 years. From 2011 to 2022, the algorithm issued four alarms: in 2013, 2017, 2019, and 2022. The alarms in 2013 and 2017 correctly predicted the events of 2014–2016 and 2018–2019, and the 2022 alarm predicted the event that began in May 2023. The only false alarm was in 2019. In the other eight years, the algorithm correctly predicted the absence of a new event. The probability of obtaining the same or better forecast configuration by random guessing was about 0.005, so randomness is virtually ruled out.
Thus, the ocean's pulse becomes an early warning. Farmers in Peru can choose rain-resistant crop varieties in advance, and water authorities in Australia can prepare for drought. The method is not perfect: one false alarm in 12 years is already too many for decision-makers. But for the first time, a long-term El Niño forecast has passed the test of reality, not just hindsight. And if we add a second step—canceling the alarm if the warm-phase index is not confirmed by the end of the year—false alarms would disappear altogether. The pulse would become even clearer.
The threshold Θ=2.82 was chosen from the range 2.815–2.834, which gave the best results on the training period 1950–1980. Interestingly, all false alarms (1994, 2004, 2019) were accompanied by at least one ONI value ≥0.5 in the same year, which inspired the idea for improving the algorithm.