Heat: how these numbers are made

The heat channel tracks two things in forty-one European cities: hot nights and days that pass each city's own extreme-heat thresholds. Every figure is one thermometer in one city, compared against that same thermometer's own history.

This page is written for someone deciding whether to cite us. It states what the numbers are, what they are not, and what has changed since we started publishing them.

What is measured

Tropical night. A calendar day whose minimum air temperature is 20.0 °C (68 °F) or above. This is the ETCCDI index TR, the standard European met services publish. We did not choose the threshold.

Warm night, the locally calibrated alternative. A night reaching this station's own 90th, 95th or 99th percentile of July-August minima over 1971-2000. This exists because TR is a Mediterranean instrument. Hamburg recorded one tropical night in the whole of 2026 and Berlin three, so a TR-based ratio there divides by almost nothing. Both metrics are published for every city; neither replaces the other, and the page states which one it is quoting.

Hot day. A day whose maximum reaches this station's own 90th, 95th or 99th percentile of July and August maxima over 1971-2000. This is AEMET's published rule, not one we invented, and it is locally calibrated by construction: 41.2 °C in Seville, 30.2 °C in Nice. A flat threshold cannot work across this range. Measured over 2011-2025, a single 35 °C bar gives 0.5 days a year in Barcelona and 66.8 in Seville.

Both, together, deliberately. Marseille is 7th of 77 years for hot nights in 2026 and 1st of 77 for hot days: the widest gap in the set between the two readings. Neither number is wrong and neither alone describes the summer, which is why the channel carries both instruments rather than combining them into one.

Sources

One station per city, named in the data. Nights and days come from the same rows of the same station record, never assembled from two.

How a year is counted

Every year is cut to the same calendar day. 2026 is compared against prior years counted to that same date, so a part-season is never ranked against complete ones. Cities in different countries have different cuts because the seven met services publish on different lags, which is why a cross-city ranking is not available and we do not print one.

A year is usable if it holds 90% of its days from 1 May to the cut. The window is measured rather than assumed: across every city and year, 99.98% of tropical nights fall on or after 1 May. Days outside it cannot hide a tropical night, so counting them toward completeness would discard usable evidence.

A tie is not a record. A year's rank counts prior years at or above it, so a tied year keeps 2026 off first place. Three cities tie 2026; resolving it the other way would manufacture a record.

A gap is not an end, and neither is a thin year. Each series carries how many slots the station record should hold, how many are present, how many were observed but too thin to rank, and how many are truly empty. Barcelona's record has real holes at 1928-1937 and 1939-1943.

Standing limits

None of these is a defect awaiting a fix. They are properties of the measurement and they are permanent.

These cities were chosen to show where the abnormal heat is, and they are not a sample of Europe. Both halves matter. This channel exists to show where the extremes are and how hard, so the cities are selected toward that and not drawn at random. The first fifteen were Spanish and French because the tropical-night measure only works where tropical nights are common; Germany and Austria reached further north; Stockholm was added to include a city that was not extreme.

The consequence, and it is the only one: no proportion here generalises to the continent. Eighteen of the forty-one are Iberia and France, there is no city east of 24.9 degrees, and none from Italy, Greece, Portugal or the UK. A set chosen at random would show a smaller share of cities at a record, because it would include places where nothing is happening.

So every count is a count of these cities. "Twenty-two of forty-one" is exactly that and never a statement about Europe.

This is not a climate measurement. A city thermometer records the city: buildings, tarmac and traffic warm it alongside any regional change. We make no attempt to separate the two, and no figure here should be read as a climate trend for the surrounding region.

One station is not a city. Madrid Retiro is a park in central Madrid. Barcelona's is the airport. A different station in the same city would give different numbers, and the choice is visible in the data rather than buried.

The 20 °C night count does not work everywhere, and we say where. 24 of the 41 cities average too few tropical nights for the measure to carry a ratio: Aberdeen, Amsterdam, Basel, Belfast, Berlin, Bilbao, Cologne, Dresden, Frankfurt, Geneva, Hamburg, Hanover, Helsinki, Leipzig, London, Munich, Nottingham, Paris, Prague, Stockholm, Strasbourg, Stuttgart, Tallinn, Zurich. Stockholm recorded none at all in 2026 and Amsterdam three. Two of them are long-standing members of the set rather than northern additions: Paris averages 1.53 tropical nights a year and Bilbao 1.17. In those cities we publish the percentile night metric instead, and the count of records we quote excludes them.

The French thresholds have weaker standing than the Spanish ones. AEMET publishes its percentile rule and we reproduce its published Madrid (36.4 °C) and Seville (41.2 °C) figures exactly. Météo-France publishes no equivalent, so the French thresholds are AEMET's method applied to French stations. Defensible, and not the same thing as citing a national service's own published number.

Three cities publish a hot-day count and no multiple. Lyon, Murcia and Palma have stations that opened in 1975, 1984 and 1978. Their 1961-1990 baselines are part-length and drawn from the warmer end of the period, so a multiple against them would understate itself while looking like the figures beside it. The number is not printed rather than printed with a caveat.

Records are rare, and a count of them needs its baseline. In a typical year almost none of these cities sets a record. Both baselines are recomputed from the series on every run rather than stored, because a stale baseline is worse than none: it looks checked.

We do not describe an elevated year as ordinary. Most of the least extreme readings here sit in the high 80s and 90s of their own records: those are elevated years that merely are not the most extreme, and calling them ordinary would misdescribe them in the direction that makes our set look tidier.

But some cities genuinely are having an unremarkable summer, and we say so. Stockholm is at the 76th percentile of its own day record. A set in which everything is extreme invites the question of how the set was chosen; the cities that are not extreme are the ones that answer it.

We also avoid universal claims about the set. Two have already gone false here as cities were added, each true when written and each really a claim about how many cities we held. Counts are used instead, because a count survives the set changing and can be checked against the table beside it.

Short records are short. Murcia's is 43 years, Palma's 49, Lyon's 52. A rank of 1 of 43 is a weaker statement than 1 of 106 and the denominator is always printed with it.

Version history

v1.3, 2026-08-08

What changed. Three cities added: Amsterdam, Stockholm and Prague, taking the set to twenty-four across seven met services. A selection block was added stating how the cities were chosen and what that forbids. Universal claims were removed in favour of counts.

What went wrong to prompt it. Two things, and they are the same thing at different scales.

Claims kept going false as the set grew. "Not one of these cities is having an ordinary summer" was true of fifteen cities and false once Berlin joined. "Every city in the set is elevated on days" was true of twenty-two and false once Stockholm joined at the 76th percentile. Both were true when written, neither produced any error when it stopped being true, and the second carried the words "verified, not asserted" directly above a hardcoded verification.

And the set was never a sample of Europe. The first fifteen cities were Spanish and French because the tropical-night measure only works where tropical nights are common, which is choosing cities on the basis of the thing being measured. That bias runs toward heat and had not been written down anywhere a reader could see it.

What it left open. The set still holds no city east of 18.1 degrees and none from Italy, Greece, Portugal or the UK, so the geographic pattern it shows cannot be separated from the geography of where we happen to measure. Whether the percentile night metric should eventually replace the 20 °C count rather than sit beside it is still undecided.

v1.2, 2026-08-07

What changed. Six cities added: Vienna, and Berlin, Hamburg, Frankfurt, Munich and Cologne. Twenty-one in total, across four national met services, all permitting commercial reuse. A percentile-based warm night metric was added alongside the 20 °C count. The headline stopped being a record count and became the geographic pattern.

What went wrong to prompt it. Two things, and the second is worse.

The 20 °C tropical-night threshold cannot reach northern Europe: Hamburg recorded one such night in 2026. Extending the city list without extending the metric would have published ratios computed against approximately zero. Checking that also revealed the same problem inside the existing set, where Paris and Bilbao average 1.53 and 1.17 tropical nights a year and had been carrying record claims on that basis.

And the lead sentence had quietly become false. It read "not one of these cities is having an ordinary summer for hot nights", which was true of fifteen cities and stopped being true when Berlin joined at the 70.9th percentile. No check covers this, because it is a true sentence about a membership that changed. It was found while auditing the word "ordinary", not by any guard.

What it left open. The geographic pattern is stated and not explained: the extreme is concentrated between roughly 38N and 51N while both ends of the range are less extreme, and we do not offer a mechanism for that. Whether the percentile night metric should eventually replace the 20 °C count rather than sit beside it is undecided. Seven cities publish a night count we do not include in the headline.

v1.1, 2026-08-07

What changed. Every city's history moved onto the national met service that the payload already named as its source. Murcia moved to the city station. Completeness is now measured over 1 May to the cut at 90%, rather than over the whole calendar year. Series carry explicit slot counts, and the tie rule is stated in the data rather than implied by the code.

What went wrong to prompt it. The payload named one source and read another, in both countries. Every Spanish city ranked an AEMET 2026 value against a history from ECA&D, a research archive that cannot be published commercially; every French city named Météo-France and read ECA&D too. It survived because the current-year value genuinely did come from the named source, so every check of recent data passed.

Murcia was worse: it was not one station. ECA&D publishes that series under the name of an air base 10 km outside the city while splicing the city station into recent decades. Our pairing matched by name, so the history came from the air base and the 2026 value from the city. The published rank had compared two different thermometers. A blended record agrees perfectly with its first component until the splice, so this is invisible to any check that scores a record as a whole.

The completeness bar was also being applied over the wrong window: a whole-year rule on a series cut in early August discarded Madrid 1936, which is complete to the cut, while keeping years whose missing days fell in July.

What it left open. Murcia's correction costs 42 years of record, and the station that is correct for a page named Murcia is also the one that keeps a record for 2026; the alternative would take the channel from eight records to seven. The northern European cities still have no working night metric. The French thresholds still rest on a method borrowed from another country's service.

v1.0, 2026-08-06

What changed. First publication. Fifteen cities, tropical nights counted against each city's own record, with per-city extreme-heat thresholds derived but not yet rendered.

What went wrong to prompt it. Not applicable; this is the baseline.

What it left open. Hot days were computed for ten Spanish stations and no French ones, leaving the channel structurally lopsided. Sourcing was not yet verified against publication licences.