CPC Went Six for Six. Two Outcomes Barely Stayed in the Winning Bin.

Public ENSO coverage usually emphasises the newest forecast, while verification of earlier monthly outlooks receives much less attention. Agencies and researchers do run formal verification; it just rarely reaches readers. This is the first of what will be a monthly check here: take the outlooks NOAA's Climate Prediction Center issued earlier this year, wait until the seasons they covered are published, and see how they did.

The answer for the seasons that can be scored today is more interesting than either "the models nailed it" or "nobody saw this coming." Both have been said about this event. Neither survives contact with the numbers, and neither does a simple hit count.

What Is Actually Being Scored

CPC publishes an Official ENSO Strength Probabilities table on the second Thursday of each month, alongside the ENSO Diagnostic Discussion. It gives, for nine overlapping three-month seasons, the percent chance that the index lands in each of nine bins: four La Niña bins, a neutral bin from −0.5 to +0.5, and four El Niño bins (weak 0.5 to 1.0, moderate 1.0 to 1.5, strong 1.5 to 2.0, and very strong at or above 2.0).

One detail on that page decides everything that follows, and it is almost never carried into coverage. CPC states it plainly: "Probabilities are verified using the Relative Oceanic Niño Index (RONI)." Not the ONI. The 1991 to 2020 base period is used for the departures.

That matters because the two indices are currently far apart. For the May-June-July 2026 season, RONI came in at +0.98°C while the conventional ONI for the same season was +1.39°C. Same ocean, same region, different arithmetic. CPC's current RONI is a three-month running mean based on ERSSTv6: it takes the Niño 3.4 anomaly, subtracts the 20°N to 20°S tropical-mean SST anomaly, then rescales the result so its variance matches the original Niño 3.4 index. It measures how much warmer Niño 3.4 is than the tropics as a whole rather than how warm it is outright, and in a warming tropical ocean the two diverge. We have written about why RONI and ONI separate and about what happens when two agencies quote different numbers for the same ocean.

Scoring a RONI-verified forecast against ONI observations would make CPC look far too cool. Everything below uses RONI on both sides, taken from the same CPC file that CPC itself verifies against.

The Scorecard

Three 2026 seasons have published RONI values and were forecast in advance: March-April-May at −0.04°C, April-May-June at +0.49°C, and May-June-July at +0.98°C. The June-July-August season does not close until 31 August, so it cannot be scored yet.

All three of those values are provisional. CPC states on the RONI page that because of the high-frequency filter applied to the ERSSTv6 data, RONI values may change for up to two months after the initial real-time value is posted, and that the most recent values should be treated as estimates. April-May-June and May-June-July are both inside that window. That matters more than a footnote normally would, because two of the three sit within 0.02°C of a category boundary, which is smaller than the revision they are still exposed to. Everything below is a scorecard against provisional observations.

Those three seasons appear in the April, May and June issuances, which gives six forecast-and-outcome pairs. "Lead" below is the number of issuances ahead. Lead 0 is a season-in-progress issuance: the season had already begun, but the completed RONI value was not yet available, so it is still a forecast rather than a report. It is not a zero-notice prediction.

One structural caveat before the table. Six pairs sound like six tests. They are not. Three of them forecast the same May-June-July season and two forecast the same April-May-June season, so the six pairs resolve to three distinct outcomes. Successive issuances are not independent of one another either, since each is built partly on the observations that shaped the last. Read the table as three outcomes seen from several distances, not as six independent trials.

IAMElNino.com analysis. Six modal hits from six attempts. The mean probability CPC assigned to the bin that actually verified was 79.5%, the average of 99, 80, 56, 84, 72 and 86. On a plain categorical scorecard that is a good run, including at two issuances of lead in a period when the Pacific was changing fast.

If we stopped here the headline would be "NOAA nailed it." That would read more into the scorecard than it holds.

The Three Hundredths of a Degree That Would Have Broken It

Look at where the observed values landed inside their bins. April-May-June verified at +0.49°C against a neutral bin whose upper edge is +0.50°C. May-June-July verified at +0.98°C against a weak bin whose upper edge is +1.00°C. Both outcomes sat within two hundredths of a degree of the top of the range CPC gave them.

IAMElNino.com analysis. Rerun the same scorecard with the ocean just 0.03°C warmer in each season, so that AMJ reads +0.52°C and MJJ reads +1.01°C, and leave the forecasts untouched. The record falls from six hits to one. The two AMJ forecasts flip from neutral to weak, a bin they had given 20% and 16%. The three MJJ forecasts flip from weak to moderate, a bin they had given 5%, 10% and 11%. Only the March-April-May forecast survives.

That is not a criticism of CPC. It is a warning about categorical verification. A hit rate records which side of a line the outcome fell on, so when outcomes keep landing near the line, the hit rate is highly sensitive to small changes near category boundaries and should not be interpreted as a complete measure of forecast skill. It is worth being explicit about what a real skill assessment would require: many independent forecasts, comparison against climatology or another baseline, and probabilistic scores such as Brier, ranked probability or logarithmic scores. Six pairs covering three outcomes support none of those. The boundary experiment shows fragility. It does not apportion the result between luck and skill, and this post does not attempt to.

The provisional status of the observations cuts the same way. A revision of 0.03°C to April-May-June or May-June-July is well inside what CPC warns is still possible, and would move those outcomes across a boundary without any forecast having changed.

A Second Look, and What It Cannot Show

There is a second, softer thing in the numbers. It needs its limits stated first, because it is easy to overclaim and we nearly did.

CPC publishes probabilities for broad bins, not a continuous distribution. A forecast of 80% neutral and 20% weak says nothing about where inside those bins CPC expected the value to fall. So there is no way to recover CPC's actual expected RONI from the published table, and nothing below should be read as one.

IAMElNino.com analysis, with the method stated in full so it can be checked or rejected. What can be done is to impute a single number from each row by assigning every bin a representative value and weighting by the published probabilities. We call the result a midpoint-imputed forecast value. It is our construction, not a CPC product. The nine assigned values are:

• at or below −2.0: −2.25  ·  −2.0 to −1.5: −1.75  ·  −1.5 to −1.0: −1.25  ·  −1.0 to −0.5: −0.75
• neutral: 0.00
• +0.5 to +1.0: +0.75  ·  +1.0 to +1.5: +1.25  ·  +1.5 to +2.0: +1.75  ·  at or above +2.0: +2.25

Four disclosures come with that. CPC probabilities are published as whole percentages. RONI is published to two decimal places. The imputed values are therefore approximate, and different reasonable within-bin assumptions will change the resulting differences. The two open-ended bins carry no upper or lower limit, so the values assigned to them (−2.25 and +2.25) are the least defensible of the nine; in the six rows scored here both open bins were given 0%, so those two choices do not enter this particular arithmetic, but wherever a tail carries any probability at all the choice would matter. Multiply, sum, and subtract from the published RONI.

Under this approximation, five of the six forecasts produced values below the subsequently published RONI, with a mean signed difference of +0.28°C. The sixth, March-April-May, sits 0.03°C the other way.

The differences also narrow as the issuance gets closer to the season. For May-June-July they run +0.50°C from April, +0.32°C from May and +0.20°C from June.

This is a diagnostic, not a finding. It is not a CPC expected value, it does not establish a statistical forecast bias, and with three distinct outcomes drawn from one developing event it could not establish one even if the method were exact. It is consistent with forecasts that sat low relative to what followed. It is equally consistent with a within-bin distribution that our midpoints misrepresent. We are reporting it because it is reproducible from published numbers, and flagging it because it is the softest thing in this article.

What the Archive Does Show

Whatever the right reading of the six pairs, one thing in the archive is not in doubt: the winter probabilities have moved a long way in four months.

Each cell is the probability that RONI reaches or exceeds +2.0°C, CPC's very strong category. In April, CPC gave the October-December peak a 23% chance of getting there. By July it gave the same season 81%.

As the event developed and new observations and model guidance became available, CPC raised the very strong probabilities each month. The archive records the revision. It does not isolate how much of it came from recent observations, how much from changes in model guidance, and how much from the forecaster team's judgment, and lead time shortened throughout, which on its own tends to sharpen a distribution.

One housekeeping note, since we quoted the older figure ourselves. Our 30 July post The Ocean Has Already Decided cited 63% for November-January. That was the June issuance and it was correct on 30 July. The July issuance moved it to 75%. The June number was not wrong; it has been superseded.

What This Does Not Show

• It does not show that the peak is settled. 81% is a forecast probability, not an outcome. It also carries a 19% chance that October-December comes in below +2.0°C, and CPC's own note that a 0% cell is approximate rather than exact cuts both ways.
• It does not show that CPC will keep revising upward. Upward revisions during the growth phase of an event are not evidence about what happens near the peak.
• It does not establish a forecast bias. The +0.28°C figure depends entirely on our own within-bin assumption and covers three distinct outcomes from one event. It is a diagnostic worth watching across future months, not a demonstrated property of CPC's outlooks.
• It does not show strong impacts. CPC says so directly on the same page: strength "does not necessarily correspond with the strength of the influence or expected impact." Index magnitude is not a regional forecast, and we do not convert one into the other. See our methodology.
• It is not a skill assessment. Six pairs resolving to three outcomes, one event, no baseline comparison and no proper probabilistic score. The public archive of the strength product also currently reaches back only to April 2026, which caps how far this check can go today.

What to Watch

• 13 August: CPC ENSO Diagnostic Discussion and the August strength probabilities table. The OND and NDJ very strong cells are the ones to read first.
• Early September: RONI for June-July-August publishes and settles the zero-lead question above.
• The gap between RONI and ONI. It was 0.41°C for May-June-July (+0.98 against +1.39). If it widens, more coverage will quote ONI-based thresholds against RONI-verified probabilities, and the two will drift further apart.
• Weekly Niño 3.4 relative values, each Monday. Seasonal RONI is a three-month running mean built on ERSSTv6, so weekly OISST numbers hint at direction but do not settle a season. They are not the same measurement and should not be compared directly.
• Revisions to the three scored seasons. All three are provisional and CPC allows up to two months of change. A revision smaller than 0.03°C would be enough to move two of them into a different category.
• Whether outcomes keep landing near bin edges. Two of three did. If that continues, categorical hit rates will keep looking more decisive than they are.

The Short Version

CPC named the right category six times out of six and put an average of 79.5% behind the category that verified. That is a real result and it deserves saying plainly, because coverage of this event has leaned hard on the idea that it caught everyone out.

It is also a smaller result than the number implies. The six pairs are three distinct outcomes. Two of the three finished within 0.02°C of the ceiling of their bin, all three are provisional and still exposed to revisions larger than that margin, and a Pacific 0.03°C warmer would have turned five of the six calls into misses. Under our own bin-midpoint approximation, five of six sat below what followed by an average of +0.28°C, which is worth watching and is not evidence of a bias.

Separately, and on much firmer ground, the archive shows the very strong probability for October-December moving from 23% in April to 81% in July. Why it moved is not something four tables can settle.

We will run this again after the August issuance, against the questions recorded above, and report whichever way they land.