Every number, and where it came from.

The indices

• Weekly and monthly Niño boxes: NOAA OISST v2.1 as published by CPC, anomalies from 1991–2020. Niño-1+2 (off Peru and Ecuador), Niño-3 (east-central), Niño-3.4 (the standard box) and Niño-4 (central). Weekly since September 1981.
• ONI: CPC's Oceanic Niño Index, the three-month running mean of ERSSTv5 Niño-3.4 against centred 30-year base periods. It is the index NOAA uses to declare El Niño and La Niña, and the one the press quotes. From 1950.
• RONI: CPC's relative ONI: Niño-3.4 minus the tropical-mean (20°S–20°N) anomaly, rescaled. It removes the lift the warming climate gives every recent event. NOAA's official strength measure, and the headline index on the Now page.
• MEI-ERA5: the Forecast Network's rebuild of NOAA's MEI.v2 method on ERA5 (sea-level pressure, SST, 2 m temperature, 10 m winds; season-specific EOFs), monthly from January 1940. It reproduces MEI.v2 at r = 0.944. The build plan's gate is r ≥ 0.98 (adding outgoing longwave radiation, as MEI.v2 uses, is the likely route), so it is labelled provisional wherever it appears. The relative variant removes the tropical basin's warming first.
• MEI.v2: NOAA PSL's Multivariate ENSO Index (JRA-3Q), two-month seasons from 1979.
• SOI: CPC's standardised Tahiti minus Darwin pressure: the atmosphere's side of ENSO. Warm water volume: NOAA PMEL's heat content above the 20 °C isotherm along the equator, the recharge that leads the next phase.

The forecasts

• Copernicus multi-system: ten C3S seasonal systems (BoM, CMCC, DWD, ECCC ×2, ECMWF, JMA, Météo-France, NCEP, Met Office), their consensus and 10–90% member range, from Forecast Season's ENSO route, on an ERA5 basis.
• ECMWF SEAS5: 51 members projected onto MEI-ERA5 (celsius.earth), and the Indian Ocean Dipole.

Record watch

Each index is compared only with itself: this event's highest value since January 2026 against the highest before 2026, same source, same base. A forecast row compares a forecast peak with the observed record and is always labelled as a forecast.

U.S. Winter Odds (history layer)

• Data: NOAA NCEI nClimDiv statewide monthly temperature and precipitation, winters 1950–51 to 2025–26.
• Strong El Niño winters: ONI classes strong and very strong from the Golden Gate Weather Services table: 1957–58, 1965–66, 1972–73, 1982–83, 1987–88, 1991–92, 1997–98, 2015–16 and 2023–24 (n = 9). Today's CPC ONI puts 1987–88 at 1.49 and 2009–10 at 1.50, so each state also reports its count with the two swapped.
• Trend first: each state and window gets a straight-line trend over the 76 winters. Thirds are cut from the detrended 1991–2020 winters, and each strong winter is placed by its detrended value.
• Hit rates, not averages: "warm third in 8 of 9", with a 95% Wilson interval. The exact binomial tail against 1-in-3 is doubled because the side is chosen after looking.
• False-discovery control: Benjamini–Hochberg at q = 0.10 across the 48 states of each map. Only states that survive are stippled.
• 2026–27 odds: each strong winter's detrended value is placed on the 2026 trend level and compared with the plain 1991–2020 thirds, beside the same calculation for all 76 winters (the trend alone).

Climate divisions and Every Strong Winter

The same method runs on NOAA's 344 climate divisions (Dec–Feb, Jan–Mar, Nov–Mar). The false-discovery test then covers all 344 at once, and with 9 winters no division clears it, so read significance on the state map. The split between cold older strong winters (1965–66, 1972–73, 1987–88) and warm recent ones survives a straight-line, quadratic or 31-year running-mean trend, and no trend at all: the 1960s–80s were a run of cold U.S. winters for reasons beyond ENSO.

What is provisional

• MEI-ERA5. The five-field version (sea-level pressure, SST, air temperature, winds) reproduces MEI.v2 at r = 0.944. Rebuilt with NOAA's full MEI.v2 recipe and outgoing longwave radiation from ERA5, it reaches r = 0.998 (docs/W4_MEI_OLR.md), passing the gate. On that version the July 2026 "record" of the five-field index does not stand: 2026 so far ranks third since 1940, behind 1982–83 and 1997–98. The five-field record came from its air-temperature field. Both are shown until the production pipeline carries the validated version.
• The model layer (SEAS5, Copernicus, NMME and SFS calibrated on their 1993–2016 reforecasts and blended with history by cross-validated skill) is not in these odds yet. Target: November 5.
• The Season Book (station odds at 52 cities) opens on October 15, once its data review is done: RCC-ACIS daily data, missing days counted, each contract with one named station and one settlement source.
• Flavour uses DJF Niño-3 against Niño-4 (ERSSTv5). For this event it uses the latest week, which is noisier than a three-month mean.

Daily boxes

Daily Niño boxes come from the Forecast Network's OSTIA store (Met Office / Copernicus Marine), anomalies against a 1991–2020 day-of-year climatology smoothed over ±15 days, computed here. They differ slightly from CPC's weekly OISST values, and both are shown with their source.

The ENSO Wheel

NOAA PMEL warm water volume (above the 20 °C isotherm, 5°N–5°S, 120°E–80°W) against Niño-3.4 (OISST from 1982, ERSSTv5 before), three-month running means, from 1980. The heat store's lead over Niño-3.4 is measured within each event's window.

Trajectory analogs

The plan's distance, D = Σ e^−k/12 [MEI(t−k) − MEI(τ−k)]² over 36 months of MEI-ERA5, matched against every February since 1943 once per SEAS5 member (51 members spliced onto the observed path through February 2027). The same method, run leave-one-out on every past February, is scored against climatology and the score is published with it. Warm water volume is left out because its record starts in 1980.

Event cards

Episodes on monthly MEI-ERA5 by the ONI rule (five or more months beyond ±0.5σ). Integrated intensity is |MEI| summed over the episode. Each event's winter is the Dec–Feb with the strongest mean inside it, and its U.S. maps use the atlas method, with the 1950–2025 trend extended to the 1940s.

Trajectory Explorer

The same 36-month recency-weighted distance, computed in your browser on the path you draw (a Catmull-Rom curve through six handles from the latest observed month to February 2028). The six closest February anchors since 1943 are shown, excluding any within two years of 2028, and their Dec–Feb winters are counted by climate division. It shows what winters after similar paths looked like, not how likely the path is.

Drought and pluvial debt

For each county, NOAA nClimDiv monthly precipitation minus that calendar month's 1991–2020 normal, summed from the start month (Sep 2020, Oct 2023 or Oct 2025) to the latest month, beside NOAA's Palmer Hydrological Drought Index. Repayment compares the debt with the Nov–Mar departure of each of the 9 strong El Niño winters. nClimDiv's precipitation files mark missing months with −9.99, and those are treated as missing, never as a value.

The world, country by country

For 237 countries and 35 named regions: population-weighted monthly temperature (ERA5, 1940→) and precipitation (GPCC Full Data Monthly V2025 at 1°, 1891→, with the First Guess product for the latest months; country-months with no rain gauge are masked). ENSO years are labelled by the January of the Dec–Feb peak and seasons run from the June–August of onset to the Dec–Feb of the year after. Each event is compared with ENSO-neutral years after removing a trend fitted to the neutral years only, and counts as a hit when it lands on the expected side of their median; counts carry Wilson intervals, a sign test and Benjamini–Hochberg control across countries and seasons. The same engine reproduces Zeke Hausfather's 26-region table (The Climate Brink, 2026-09-28): 179 of 208 region-events against his 176 of 205, identical on the 22 regions where he used GPCC. Ten other modes (IOD, PDO, IPO, AMO, NAO, AO, PNA, SAM, QBO and the tropical Atlantic) are correlated at lag 0 and 1 with ENSO partialled out; decadal views use 11-year and 2–7-year filters. Details: docs/GLOBAL_DRIVERS.md.

Model agreement abroad

Ten Copernicus systems and three NMME models, one vote each, each judged against its own 1993–2016 hindcast at the same start month and lead (terciles from its members), with its hindcast correlation against ERA5 and a false-discovery test across the map. Agreement of 80% or more is flagged robust, but related models are not independent votes, so no interval is attached. A September start reaches February; spring and summer 2027 become reachable with the December and February starts. Details: docs/MODELS_WORLD.md.

The planet

A monthly regression per dataset (ERA5, NOAAGlobalTemp, HadCRUT5, GISTEMP): a quadratic trend since 1970 (chosen by a rolling 1995–2025 hindcast), Niño-3.4 at its best lag (3 months; nearly flat from 2 to 5), and a two-speed residual. 4,000 draws follow the Copernicus Niño-3.4 plume to February 2027 and the decay of a randomly chosen past strong El Niño after it. Year-ahead error 0.081 °C against 0.127 for the trend alone; the 90% intervals held 94% of the time. CO₂ growth uses NOAA GML Mauna Loa with the same lagged regression.

The year after, archetypes and the transition

The year after uses the 8 events with an Nov–Jan ONI of 1.5 or more, split by whether the next winter's lowest ONI reached −0.5. Archetypes are stated rules on three consecutive Dec–Feb winters on MEI-ERA5 (±0.5σ), with U.S. division composites for each year of the sequence and a shuffle test of the whole map. The transition watch shows warm water volume, the models and the historical record side by side and does not blend them, because the blend has no track record to score.

Sectors

Hurricanes: NHC HURDAT2 (Atlantic 1851→, eastern and central Pacific 1949→) and IBTrACS for the season in progress; ACE from 6-hourly synoptic fixes at 35 kt or more; the Atlantic's warmth is the ERA5 Main Development Region minus the tropical mean. Severe weather: SPC's database, EF1+ January–May tornadoes in eight southern states. Energy: nClimDiv population-weighted heating degree days, census divisions weighted by 2020 population; EIA working-gas storage. Agriculture: NCEI's crop regions. Water: April 1 snow water at 1,358 long-record SNOTEL and snow-course sites as an NRCS-style basin index (sites that did not report are left out, never counted as zero); CDEC station indices.

The ocean

Equatorial (5°S–5°N) daily OISST v2.1, 850 hPa wind from NCEP/NCAR Reanalysis 1 (GFS analyses adjusted to it on 441 overlapping days after Reanalysis 1 stopped updating on 2026-03-17), the NOAA OLR CDR, and GODAS temperature to 459 m, all against 1991–2020. A westerly wind burst is an 850 hPa anomaly above +5 m/s over at least 10° of longitude on at least three consecutive days.

Build

Everything is precomputed on the Forecast Network's servers from the public files named above, then published as static pages and JSON. A source that fails to update keeps its last good copy, marked stale in /data/manifest.json. Nothing is filled in by hand.