Methodology

The bright line

This site shows the observed record only. No forecasts, no projections, no scenarios. Every number on every chart describes a day, a month or a year that has already happened.

There is one place a model field is used, and it is bounded. Reanalysis — the observational record of the atmosphere — runs about five days behind the present. To carry the daily temperature charts to yesterday rather than stopping five days short, the last few days are filled with short-range model fields from ECMWF's HRES for elapsed days: the 00Z run for a day that has already finished, read after the fact. Those days are drawn dashed with a hollow endpoint and labelled provisional, and each one is replaced by reanalysis the moment reanalysis reaches it.

A provisional value is an estimate of a day that has happened. It is never a forecast of a day that has not. No forecast field is published anywhere on this site, and none ever will be. That is the editorial position the whole site is held to: a site that published model forecasts alongside observations would blur exactly the line that makes it worth reading.

Three tiers of recency

Every row of every published series carries a status, and that field drives the styling all the way through to the downloadable image. For the ERA5-based temperature charts it takes three values.

The tier is read from the data, not guessed. CDS serves ERA5 and ERA5T from one dataset and distinguishes them with an expver coordinate, one value per timestep. The status of a day is taken from that directly, never inferred from the date.

Promotion to final. Final ERA5 arrives two to three months after ERA5T, by which time a day has long left the window that is re-fetched each morning. So promotion is its own step: each morning, the oldest month that still holds ERA5T days is replaced with final ERA5 from NCAR's mirror of the archive, as soon as that mirror holds the month. The mirror carries final ERA5 only, so a month appearing on it is itself the evidence that the final version exists. When the same day is held at two tiers, the more settled one always wins.

ERA5T is not styled as provisional, deliberately. It is preliminary reanalysis, produced by the same system as ERA5 and differing only in that a few late observations may still arrive. It is not a model estimate. Marking it would spend a reader's attention on a distinction that does not change how the number should be read, and would dilute the marking that does — the HRES tail.

How the provisional days are built

The HRES conventions are fixed, because the correction described below is defined against them and a silent change to any of them would move every provisional value:

• The 00Z cycle only. Never 06, 12 or 18.
• Steps 0, 3, 6 … 21 — the eight three-hourly samples that fall inside the day. Step 24 is the next day's midnight and belongs to the next day.
• Only the 2 m temperature message is fetched, by byte range: 0.65 MB rather than the ~139 MB the full field file weighs.

Eight three-hourly samples are not twenty-four hourly ones, so an HRES daily mean and a reanalysis daily mean do not agree exactly. The difference is measured rather than assumed. On every run the two are compared across the five most recent days both of them cover, and the provisional days are corrected by the measured offset before publication. Without that correction there is a visible step exactly where the provisional tail meets the reanalysis, and drawing two sources as one line without removing it would be dishonest.

The offset is re-measured every run rather than stored. It is small — under 0.03 °C for the globe, the Arctic and the contiguous US, and about −0.26 °C for Europe — but it follows the model, and an ECMWF cycle upgrade moves it. A number frozen in configuration would go on applying the old one without saying so. Every comparison is also kept, so that a step in the offset is visible after the fact and raises an alert: the correction absorbs it either way, but a change in the relationship between model and reanalysis is something a reader of this page deserves to be able to look up.

Baselines

An anomaly is only meaningful against a stated reference period. This site uses 1991–2020 as the operational baseline everywhere, which is the current WMO normal period and the one most operational products use. Where a chart says "normal" or "anomaly" without qualification, that is the period meant.

One fixed baseline across every chart is a deliberate trade. It means the same colour means the same anomaly on every map and the same distance means the same departure on every series, which is what makes thirty charts one instrument rather than thirty. The cost is that a value here can differ from an official product defined another way, and where that matters the chart says so.

El Niño is measured relative to the whole tropical ocean. Since February 2026 NOAA classifies El Niño and La Niña with the Relative Oceanic Niño Index (RONI): the Niño 3.4 anomaly minus the anomaly of the entire tropical ocean between 20°S and 20°N, rescaled so its spread matches the plain index. The subtraction keeps the general warming of every tropical ocean from being counted as El Niño. The Niño 3.4 chart here uses the same method, rescaled over 1991–2020, with NOAA's thresholds of ±0.5 °C read against the three-month mean. It is built from OISST, where NOAA uses ERSSTv6, so the two are close but not identical. No “super” level is shown, because NOAA defines none.

Headline global temperature is the exception the spec asks for: it belongs against 1850–1900, the pre-industrial reference. ERA5 begins in 1940 and cannot produce that period, so the intended method is a 1991–2020 climatology plus a fixed published pre-industrial offset. See known limits — this is not yet what the live global charts do, and the page says so rather than implying otherwise.

What is done to the numbers

Every published series is derived from its source by one of a small number of steps, applied the same way for every indicator that uses them.

• Day-of-year climatologies average each calendar day over the baseline period. Leap years are aligned by slot, and 29 February — which has roughly a quarter of the years behind it that its neighbours have — is interpolated from them rather than published as a real value, because otherwise it appears as the largest single-day step in the year and the step is an artefact of the calendar.
• Regional means are area-weighted by the cosine of latitude, and land or ocean masked where the region says so. Every region is cut from one global field rather than fetched separately, so two regions can never disagree about a day.
• Running means are centred, run over calendar months, and are published only where the window is complete. A three-month mean of two months would move when the third month lands, so a smoothed series ends short of the monthly one — one month for a three-month mean, five for a twelve-month one — and a window that would span a missing month is not published at all.
• Anomaly maps show the mean of the last 30 days against the same 30 calendar days of the 1991–2020 baseline, on a 0.5° grid, drawn as discrete cells at the data's own resolution with no smoothing. They end where reanalysis ends: a provisional field is not drawn cell by cell, because doing so would imply a precision it does not have.
• Trends are least squares, with the 95% interval widened for lag-1 autocorrelation and taken from Student's t on the effective number of independent years. On the Dutch sea level record, ordinary least squares would understate the interval by about a quarter. An uncertainty band that is too narrow is worse than none — and one that is too wide is no better. The autocorrelation is measured about a curve, not a line: the satellite record is accelerating, and measured about a straight line, that acceleration would be read as noise, inflating the rate's interval to three times the size its producer can support. Where a producer publishes the measuring system's own error on the rate, as AVISO does for altimetry, it is added in quadrature, because no fit to a system's output can see an error the system makes the same way every year.

What happens when something is wrong

Every update runs the same checks, and they are split into two kinds. The split is the important part.

Blocking checks mean the data is wrong. A value outside physical bounds, dates that are duplicated or out of order, or a multi-day hole in the recent interior where the source does have the data — each of these means the pipeline got it wrong, and the new data is refused. The previously published series stays exactly where it was, with its own honest "updated" stamp.

Advisory checks mean the data is merely old or incomplete, which is a fact about the world rather than an error. A source running late publishes with its real date. A day the source never published leaves a break in the line and a note underneath saying which days are missing — gaps are never interpolated across, because a line drawn through a hole is a claim nobody made.

A record that has never reached the present at all is withheld rather than published, because a chart ending in 1996 under a caption promising the present misleads in a way a late chart does not. And a chart that stops advancing for more than three times its own allowance raises an alert, since from the outside a frozen chart and a chart with nothing new to say look identical.

Sources, attribution and licence

Each chart carries its source and licence in its own signature block, so the attribution travels with every screenshot and every embed. Those lines are generated from the source register, never typed by hand, which is why they cannot drift from what was actually used.

The derived series and the charts are published under CC-BY 4.0. The code is MIT. Each upstream source carries its own licence and its own required attribution, listed per indicator below.

Known limits

Things that are true about this site today and that a careful reader would otherwise have to discover for themselves.

• The headline global temperature chart is not yet a pre-industrial anomaly. It publishes absolute daily global mean temperature against the 1991–2020 normal. The 1850–1900 version the spec describes needs the offset method in baselines and an anomaly mode for the chart; until both exist, the chart is honest about what it shows but it is not the headline number.
• The global monthly long record is not published at all. Its baseline is 1850–1900 and its primary source begins in 1940, so it cannot be built as configured. It is listed below as pending rather than quietly omitted.
• 9 of the 30 charts are not live yet. Each is marked below.
• Provisional days are corrected, and a correction is an estimate. The offset is measured over five days, so it removes the systematic part of the difference and not the day-to-day scatter. A provisional value is a good estimate of an elapsed day, not a measurement of it. That is what the dashed styling is for.

Every indicator

Generated from the frozen indicator list, so it describes what the pipeline actually does rather than what someone remembered to write down. Each chart links here, to its own entry.

Arctic

Global

Tropics

Europe

Contiguous US

Netherlands & sea level