Forecasting the El Niño–Southern Oscillation (ENSO) has remained a challenge for decades because of its global socioeconomic impacts. Linear inverse models (LIMs) are simple and work well on short time scales, but their linear structure fails to capture nonlinear interactions, especially during El Niño onset. Around this work in the archive lie studies on predicting El Niño a year and a half ahead, on tracking indoor heat with a single sensor, on algorithms that anticipate chaos, and on forecasting solar flares with confidence. All of them, like this paper, seek ways to extract predictability from complex dynamics, but this work stands out by applying Koopman operator theory to climate data.
The authors build forecasts of the Niño 3.4 index (the average sea surface temperature anomaly in the central Pacific) from historical sea surface temperature fields in the Indo-Pacific region. Instead of linear regression, they use kernel analog forecasting (KAF): the method looks for past states similar to the current one and averages their future evolution, but it does so in a nonlinear feature space constructed with kernel functions. A key detail is lags: predictors include not just current temperature snapshots but sequences of 12 months, which helps account for hidden variables such as upper-ocean heat content. Training used HadISST data from 1870 to 1997, validation from 1998 to 2017, and also a 1300-year control run of the CCSM4 climate model.
On observational data, KAF extended the useful forecast horizon (correlation coefficient above 0.5) from 6–7 months for LIM to 9–14 months depending on the decade. On average over the 20 years, the gain was 3–4 months, and in some periods it reached 7 months. On CCSM4 model data, KAF's advantage was stronger: a correlation of 0.5 persisted up to 24 months versus 18 months for LIM. The method also softened the "spring predictability barrier," the traditional drop in forecast accuracy when crossing spring: in June–September, KAF's correlation remained around 0.5–0.6, whereas LIM's fell to 0–0.3. For the historical El Niño of 2015/16, a forecast initialized in June 2015 predicted a strong event with a peak in December, underestimating the amplitude by only 0.4°C. Probabilistic forecasts of El Niño and La Niña events were also more accurate according to Brier score and relative entropy metrics.
The work shows that nonparametric methods based on Koopman operator theory can outperform linear models in ENSO forecasting without resorting to expensive dynamical models. Unlike neighboring works that either use linear approximations or require huge datasets to train neural networks, KAF learns from a relatively short observational record (~120 years) and yields a theoretically optimal forecast as a conditional expectation. This opens a path to improving seasonal climate forecasts at minimal computational cost.
The method can be extended to other geophysical phenomena with nonlinear dynamics, such as forecasting monsoons or North Atlantic oscillations. KAF could also be combined with ensemble approaches to estimate uncertainty. Including additional variables, such as wind or subsurface temperature, is also promising for further extending the forecast horizon.
Better El Niño forecasts will directly affect agriculture, water resource management, energy, and insurance in ENSO-sensitive regions. Probabilistic forecasts will enable more informed decisions about planting areas, fuel reserves, and disaster preparedness.
Test the method on other datasets, including higher-resolution reanalyses, and investigate sensitivity to kernel parameter choices and lag length. It is also worth comparing KAF with state-of-the-art dynamical models in an operational setting.
In climate forecasting, the problem of predictability beyond a few months remains unsolved because of the chaotic nature of the atmosphere. This work shows that nonlinear statistical methods can partially overcome this barrier, but the question of the physical limits of ENSO predictability remains open.
The 2015/16 El Niño was so strong that it raised global temperatures by 0.1°C, and coral reefs worldwide experienced the third global bleaching event.
In Kim Stanley Robinson's novel "Forty Signs of Rain," climatologists struggle with the consequences of an El Niño amplified by climate change. The KAF method could have given them earlier warning of the impending disaster.