One independent synthesis of 14 dynamical models and 667 ensemble members points toward a record-scale peak. Here is how El Niño prediction actually works: dynamical versus statistical, why the models disagree, which ones have earned their track record, and what a multi-model plume can and cannot promise.
Every "record El Niño" headline this summer traces back to the same place: the output of computer models. When NOAA assigns an 81 percent chance of a very strong event by winter, that number is not a hunch, but it is not a mechanical model average either. It reflects an expert synthesis of multi-model forecasts, current observations, reanalysis and the evolving state of the coupled ocean-atmosphere system. This post is about what actually sits behind those probabilities: the two families of El Niño model, the historic systems that started it all, which ones forecasters trust most, and the honest limits of the whole enterprise.
Essentially every ENSO forecast comes from one of two approaches, or a blend of both.
Dynamical models numerically approximate the climate system using physical equations and parameterizations. They take the observed state of the ocean and atmosphere, encode the equations that govern heat, wind, pressure and currents, and step the simulated system forward in short numerical time increments. These range from simplified models of just the tropical Pacific to full coupled general circulation models that simulate the entire globe.
Statistical models take the opposite route. Instead of simulating physics, they learn from history: they mine decades of observations for patterns that reliably precede an El Niño or La Niña, then project those patterns forward. The methods span simple linear regression, canonical correlation analysis, and, increasingly, machine learning and deep neural networks. According to the International Research Institute for Climate and Society, dynamical models tend to do better at catching the onset of an event in boreal spring and summer, while statistical models are competitive at tracking an event once it is underway through fall and winter.
A modern dynamical forecast has three moving parts. First, initialization: the model has to start from where the real world is right now, so observed sea surface temperatures, subsurface ocean heat, winds and pressure are blended into the model's starting state through a process called data assimilation. Small errors here can grow, which is a big reason forecasts get harder the further out you look.
Second, coupling. In coupled dynamical ENSO models, ocean-atmosphere feedbacks are essential. A warm ocean weakens the trade winds, and weaker winds warm the ocean further. A model that does not couple those two systems tightly will miss the feedback that defines El Niño.
Third, the ensemble. No single run is trusted on its own. Forecasters nudge the starting conditions in dozens of slightly different ways and run the model many times. If most of those runs cluster together, the internal ensemble uncertainty is smaller, though common model biases can still make the forecast wrong. If they scatter, the future is genuinely uncertain, and the model is honest enough to say so. That spread is what you see when a forecast is drawn as a plume.
Statistical ENSO prediction has deep roots and remains operationally relevant. Canonical correlation analysis, applied to ENSO prediction by Barnston and Ropelewski in 1992, identifies sequences of predictor patterns that tend to evolve into a forecastable outcome, and versions of it have run operationally at NOAA for decades. It is cheap, transparent, and surprisingly hard to beat at some lead times.
The newer entrants are machine learning models: convolutional networks, recurrent and long short-term memory networks, transformers and graph neural networks trained on climate model output and observations. A growing body of research reports that these can match or exceed individual dynamical models at longer leads, and the most promising recent work combines the two, using deep learning to correct the biases of a physics model. None of this replaces the physics. It supplements it.
If one model deserves a plaque, it is the Zebiak-Cane model. Developed by Mark Cane and Stephen Zebiak at Columbia University's Lamont-Doherty Earth Observatory, it produced the first successful real-time dynamical ENSO forecast in 1986, with Sean Dolan joining Cane and Zebiak on the published experimental forecast. It demonstrated that coupled dynamical ENSO prediction was possible and helped establish the conceptual foundation for later operational forecasting systems. It remains one of the most cited and studied models in the field.
Two big multi-model efforts anchor operational ENSO forecasting. The North American Multi-Model Ensemble (NMME) has pooled systems from NOAA, NASA, GFDL, NCAR and Environment and Climate Change Canada, among them CFSv2, GFDL's FLOR, GEOS-5, the Canadian CanCM models and NCAR's CESM, with the precise operational roster and model versions changing over time. Across the Atlantic, the Copernicus Climate Change Service (C3S) pools the European systems: the ECMWF, the UK Met Office, Meteo-France, Germany's DWD, Italy's CMCC and others. NMME and C3S are the two major multi-model systems. The IRI separately assembles a broad international ENSO plume, while NOAA forecasters weigh model guidance alongside observations when issuing the official outlook.
On track record, the European Centre's SEAS5 system is widely regarded as one of the most skillful, with Niño 3.4 correlations reported near 0.9 at short lead times in at least one model comparison, though skill falls off as lead time grows. NCEP's CFSv2 is a workhorse but carries known biases, including a tendency toward larger tropical SST variability that inflates its errors. The honest bottom line, established across a decade of NMME research, is that no single model is reliably best. The multi-model average has often beaten the individual systems, which is exactly why forecasters quote the ensemble rather than any one model's headline number.
ENSO forecasts have a famous weak spot called the spring predictability barrier. Forecasts launched in February through May lose skill sharply, because spring is when the tropical Pacific's temperature gradient is weakest, ocean-atmosphere coupling is loosest, and the anomalies a model is trying to detect are smallest against the background noise. A forecast made in April is simply working with less signal than one made in September.
That is what makes 2026 unusual. This event launched from near-La Niña conditions in January, when NOAA had the weekly Niño 3.4 index at -0.9°C in February, and strengthened straight through the spring barrier to +1.2°C by early July. One independent analysis by climate scientist Zeke Hausfather is worth reading carefully here, because it is a custom synthesis, not NOAA's official product. In that analysis, the model-weighted median of the forecast peak (measured as the monthly, detrended Niño 3.4 anomaly, which is not identical to NOAA's seasonal RONI strength categories) rose from roughly 2.8°C in the March synthesis to 3.6°C in July. The upward revisions have recently become smaller, but more initializations are needed before calling the forecast stable. The spread is also real: at the July initialization one model (CMCC) sat far above the pack near 5.3°C while another (JAMSTEC's SINTEX-F) was the low outlier around 2.2°C. Discount the extremes, watch the middle.
A forecast model is a disciplined way of being uncertain, not a crystal ball. It can tell you the odds, quantify how much the credible outcomes disagree, and update honestly as new observations arrive. It cannot promise a specific number five months out, and it cannot verify itself in conditions it has never encountered. That is not a knock on the science. It is the science being honest about its own limits, which is the part the headlines usually leave out.
The next scheduled model-driven update is NOAA's ENSO Diagnostic Discussion on August 13. You can watch the live indices that feed these systems on our teleconnections dashboard.