Inside NOAA's Crystal Ball: How Scientists Predict El Niño

Published: May 22, 2026 · 10 min read

In May 2026, NOAA issued an El Niño Watch with an 82% probability of onset by July. At the same time, forecasts for peak intensity ranged from "moderate" to "super El Niño exceeding +3°C" — depending on which model you asked. The ECMWF's European model said 100% probability of a super event by November. NOAA's own ensemble gave a 37% chance. A third model cluster suggested the whole thing might fizzle.

How can the world's best climate models disagree this much about something that's already beginning to happen?

Answer: the spring predictability barrier — one of the most stubborn problems in climate science, and the reason why ENSO forecasts made between March and May carry more uncertainty than those made at any other time of year.

The Three Layers of ENSO Prediction

Predicting El Niño isn't like predicting tomorrow's weather. Weather models look 7-10 days ahead. ENSO forecasts look 6-12 months ahead. The difference in timescale means completely different prediction approaches.

Layer 1: Direct Observation

The Tropical Atmosphere Ocean (TAO) array — 55 moored buoys strung across the equatorial Pacific — measures wind speed, air temperature, humidity, and ocean temperature at multiple depths. These buoys are the ground truth. They tell us what the ocean is doing right now: the current Niño 3.4 anomaly (+0.4°C as of mid-May 2026), subsurface temperatures (+1.6°C at 300m), and wind patterns.

Observation tells you where you are. It doesn't tell you where you're going.

Layer 2: Dynamical Models

These are the heavy hitters — full-physics climate models that simulate the coupled ocean-atmosphere system. NOAA's CFSv2 (Climate Forecast System version 2), ECMWF's SEAS5, and Japan's JMA model all fall into this category. They ingest current observations, then run physics equations forward in time to project SST anomalies months ahead.

Dynamical models are powerful but hungry. They require enormous computing resources. They're also sensitive to initial conditions — if you feed one slightly different starting data, you can get substantially different outcomes 6 months out.

Layer 3: Statistical Models

These are simpler: feed in historical patterns, find the events that looked most like what's happening now, project forward based on what happened next in those historical analogs. Statistical models are computationally cheap and often competitive with dynamical models for short-range ENSO forecasts. But they struggle with unprecedented situations — and a world where every El Niño lands on a warmer baseline is, by definition, unprecedented.

The North American Multi-Model Ensemble (NMME)

NOAA doesn't rely on any single model. The NMME combines output from multiple dynamical and statistical models — currently 8+ members — into a single ensemble forecast. When you see "82% probability of El Niño," that's not one model's opinion. It's the fraction of NMME ensemble members that cross the El Niño threshold.

The ensemble approach smooths out individual model biases. But it also masks disagreement. When the ECMWF model says "super El Niño certain" and another model says "barely El Niño," the ensemble might show "moderate El Niño probable" — which is neither model's actual forecast. Understanding the spread matters as much as the central estimate.

In May 2026, the NMME ensemble mean approaches +1.5°C by May-July. But the spread is enormous: individual members range from +0.5°C to +3.0°C+. That spread is the spring predictability barrier in action.

The Spring Predictability Barrier: Why May Forecasts Are the Least Reliable

The spring barrier isn't a theory. It's a statistical fact, documented in decades of forecast verification studies.

The problem is physical: during boreal spring (March-May), the equatorial Pacific is in a transitional state. Trade winds are at their weakest. The ocean-atmosphere coupling that drives ENSO — the feedback loop between SST anomalies and wind patterns — is at its lowest signal-to-noise ratio. Small perturbations in initial conditions get amplified in the models because the background coupling is weak.

Practical consequence: a forecast issued in May has roughly half the skill of a forecast issued in August for predicting December conditions. The models don't get smarter in August. The signal just gets stronger.

A 2025 paper in npj Climate and Atmospheric Science examined why NMME models generate "spring false alarms" — predicting El Niños that never materialize. The culprit: models overestimate the persistence of springtime westerly wind anomalies. When the real atmosphere doesn't sustain those winds, the model's predicted warming doesn't happen. But the model keeps predicting it until new observations overwhelm the old trajectory.

Kelvin Waves: The Subsurface Telegraph

Before sea surface temperatures in the eastern Pacific warm up, Kelvin waves do the work underwater. These are eastward-propagating waves that travel along the thermocline — the boundary between warm surface water and cold deep water — at depths of 100-250 meters.

A downwelling Kelvin wave depresses the thermocline: warm water piles up on top, preventing cold upwelling. When the wave reaches South America (typically 2-3 months after forming in the western Pacific), the trapped heat surfaces.

This is why subsurface temperature is such a critical leading indicator. The Niño 3.4 surface temperature is +0.4°C in May 2026. But the Kelvin wave train already in transit carries water that's 5°C above normal at 150m depth. When that water surfaces — which it will, in roughly June-July — the Niño 3.4 index will jump.

Four significant Kelvin wave events have been detected since December 2025. The most recent, forming in April 2026, is described by NOAA's Michelle L'Heureux as "rivaling the one we saw in 1997." That's not a guarantee of a 1997-magnitude event. But it's the strongest subsurface signal in nearly 30 years.

How to Read an ENSO Forecast (Without Being a Climate Scientist)

When you see an ENSO forecast headline, check three things:

1. When was it issued? A May forecast carries roughly 2x the uncertainty of an August forecast. If it's March-May, mentally widen the error bars.

2. What's the ensemble spread? "60% probability" on an ensemble with tight agreement is very different from "60% probability" where 6 models say yes and 4 say no. If the spread isn't mentioned, the forecast is incomplete.

3. What's happening subsurface? Surface temperatures can wobble by 0.2-0.3°C from week to week. Subsurface heat content moves more slowly and is harder to reverse. If subsurface heat is rising for 6 consecutive months (as it is in early 2026), the signal is real regardless of weekly surface wobbles.

The 2026 El Niño forecast will become much clearer by July. Until then, the data says: El Niño is coming. The models disagree on how big. The smart bet is to prepare for a strong event and hope the conservative models are right.

Data sources: NOAA CPC ENSO Diagnostic Discussion (May 14, 2026); "Understanding spring forecast El Niño false alarms in the NMME" — npj Climate and Atmospheric Science (2025); Climate Impact Company ENSO outlook (January 2026); Climate Cosmos subsurface analysis (May 8, 2026); TAO/TRITON array data (NOAA PMEL).