A February 2026 arXiv preprint gave El Niño a 37.5% chance this year and put the climate-network method at 91.4% for no onset at all. The event arrived anyway. But the same paper called the Eastern Pacific flavour correctly, and the combined probability rested on a sample of six.
On 16 February 2026, Ludescher, Meng, Fan, Bunde and Schellnhuber posted an arXiv preprint giving their experimental long-lead ENSO forecast for the year. It is a v1 preprint, not a peer-reviewed paper, and we will keep saying so. But it is dated, public and specific, which makes it exactly the kind of forecast that can be scored.
Their climate-network method put the probability of no El Niño onset in 2026 at 91.4%. Their combined forecast gave El Niño 37.5%, below NOAA CPC's 61% for August to October, issued the month before and cited in the paper itself.
El Niño developed. CPC now gives a greater than 90% chance of a very strong event, and a 75% chance that October to December exceeds every event since 1950.
The easy story is that the forecast failed. The accurate story is more useful, because roughly half of it worked.
| Component | Forecast | Verdict |
|---|---|---|
| Climate network, onset | 91.4% no El Niño | Wrong |
| Complexity, onset | 71.4% El Niño | Effectively right |
| Complexity, magnitude | 0.84 ± 0.36°C maximum ONI | Already badly underpredicted; final error awaits peak ONI |
| Type, EP or CP | 76.9% Eastern Pacific | Right |
| Combined onset | 37.5% El Niño | Wrong |
| NOAA CPC, January 2026 | 61% El Niño for ASO | Right |
Two of the components landed. The complexity-based method called the onset at 71.4%, and the separate type forecast called an Eastern Pacific event at 76.9%. The developing event is strongly east-weighted, with Niño 1+2 at +3.9°C against Niño 4 at 0.0°C in the week centred 16 September. That is a real success and it should not be buried.
What failed was the network method, the magnitude estimate, and the rule used to reconcile the two.
The complexity method forecasts the maximum traditional ONI of the event, at 0.84 ± 0.36°C. Three standard deviations above that central value is +1.92°C.
CPC's ONI for June to August 2026 is already +1.80°C. That is 2.67 standard deviations above the forecast, and the event is still strengthening. If the traditional ONI reaches +2.0°C, the forecast lands more than three standard deviations low.
We are not calling this component finally wrong, because the quantity forecast was the peak and the peak has not happened. But the direction and rough size of the error are already clear.
This is the part worth slowing down for, and the authors deserve credit for stating it openly rather than hiding it in a supplement.
The two methods disagreed. The network said no, the complexity method said yes. So how do you combine them into one number? The authors went and counted what happened the previous times they disagreed:
"In the hindcasted and forecasted past (1984-2025), there were 6 cases where an El Niño onset prediction by the complexity-based method was not matched by the climate network-based method... In 2 of these cases, an El Niño did start (2004, 2006) and in 4 cases, it did not start (1985, 1993, 2012, 2022)."
Applying Laplace's rule of succession to those six cases gives (2+1)/(6+2) = 37.5%.
Six cases. That is the entire evidential basis for the headline combined probability. Laplace's rule is a reasonable choice for a small sample, and using it is more honest than inventing a weighting scheme. But a probability built on six observations should be read as a rough prior, not a calibrated forecast, and the gap between 37.5% and NOAA's 61% is smaller than it looks once you know the denominator.
2026 is now the seventh case in which these two methods disagreed, and this time El Niño started. Running the authors' own procedure with the new case included:
(3 + 1) / (7 + 2) = 4/9 = 44.4%
Adding 2026 as the seventh disagreement case changes the rule-of-succession estimate from 37.5% to 44.4% for a future case in which these two methods disagree. That is a mechanical retrospective update, not a newly validated forecast, and it is what the authors' own method produces when the new data point is fed back in.
Two things prevent us from closing the book, and both matter more than they look.
The qualifying-season count is ambiguous. The paper scores itself against traditional ONI, whose retrospective definition requires five consecutive overlapping seasons at or above +0.5°C. CPC's published grid rounds MAM 2026 to +0.5, although the underlying file gives 0.46. CPC does not specify whether the five-season threshold is applied before or after rounding. Under the published-grid convention, JAS would become the fifth consecutive qualifying season if it remains at or above +0.5. Under the unrounded convention, the sequence begins with AMJ and ASO would be the fifth. Either way, the paper's traditional-ONI verification is not yet final, even though NOAA has already declared El Niño using RONI.
The index version has changed underneath the forecast. The paper refers to the 2025 version of traditional ONI. CPC's current grid is ERSSTv6. A final score should be evaluated using the same index version and classification convention the authors used, or the difference should be disclosed. We are disclosing it.
| Season | ONI (file) | Grid | Qualifies? |
|---|---|---|---|
| MAM 2026 | 0.46 | +0.5 | Depends on rounding convention |
| AMJ 2026 | 0.95 | +1.0 | Yes |
| MJJ 2026 | 1.39 | +1.4 | Yes |
| JJA 2026 | 1.80 | +1.8 | Yes |
The track record here is genuinely strong, which is what makes the case instructive rather than embarrassing.
The authors report 14 real-time forecasts for 2012 to 2025, of which 13 were correct, the single error a false alarm in September 2019. A more restrictive version of the algorithm gave nine alarms, all correct, with no false alarm. That is a record most forecasters would be pleased with.
But the same restrictive version missed three El Niño onsets. A method that never cries wolf and sometimes sleeps through the wolf has a particular failure signature, and 2026 fits it exactly. The miss does not expose fraud or invalidate the approach. It exposes the difference between a method that reconstructs the past well and a method you can bet on in real time.
As for why, the authors offer a candidate. They identified the different reanalysis inputs, NCEP Reanalysis 1 for the network method and ERA5 for the complexity method, as a possible reason the two methods disagreed. Whether that also explains why the NCEP-based network method missed the event remains an open question. We covered the broader difficulty in how ENSO forecast models work, and the spring predictability barrier is the reason any forecast issued in February deserves wide error bars.
Closes the magnitude score: the peak traditional ONI for this event. If it reaches +2.0°C the forecast of 0.84 ± 0.36 lands more than three standard deviations low. If the event stalls near +1.2°C it lands close to one standard deviation high on the low side, a very different verdict.
Closes the onset score: JAS or ASO 2026 completing five consecutive qualifying ONI seasons, depending on the rounding convention applied.
Would complicate this piece: a revision to recent ONI values. CPC warns that recent seasons remain subject to revision, and a downward revision to MAM or AMJ would move the qualifying sequence again.
We will re-run it: when the group publishes its 2027 forecast, we will score this one to completion and check whether the combined rule was updated.
The experiment did not simply fail. One method detected the onset and the separate type forecast correctly identified an eastern-Pacific event. The network method missed the onset, the magnitude estimate badly undershot the developing event, and the rule used to reconcile the disagreement favoured the wrong outcome. That split result is more scientifically useful than either declaring victory or dismissing the entire approach.