10 public forecast models for London's overnight low, logged every evening and checked every morning. Nothing dropped, nothing chosen afterwards.
Mean absolute error against the settled whole degree. Average of all 10: 0.52 °C. Best single model, AROME France HD: 0.55 °C. 42 nights, 4 October – 14 November 2026.
The same errors, signed. GEM (Canada) runs 1.75 °C low, on 38 of 42 nights; UKMO 2 km (Met Office) runs 0.78 °C high. A constant offset is the most fixable kind of wrong.
Overnight low vs. daytime high. AROME France HD is among the best at night and near the bottom by day; AROME France HD leads the day. Skill does not carry over. 42 nights each.
Each night's error for the 10-model average. Inside the band it rounds to the settled degree: 25 of 42 nights. Vertical position is spacing only.
All 1023 combinations of the 10 models, averaged. One model misses by 0.92 °C; any two, 0.72 °C. The line is each row's average. 42 nights.
Same test on the daytime high: average of 10 0.71 °C, AROME France HD 0.59 °C. Errors agree more by day (r = 0.53) than at night (r = 0.36), so averaging cancels less. 42 nights each.
42 nights resampled with replacement, 20,000 times. The average of 10 led in 62%; 95% put its lead between -0.16 and +0.22 °C. That range crosses zero: a real lead, not yet a settled fact.
Open-Meteo: ECMWF, GFS, ICON, UKMO, AROME, GEM, JMA and two HARMONIE runs. Read at the same moment each evening, stored before the outcome exists.
Scored against the whole degree a public market settled on. External, published, and beyond our reach to revise.
No filtering, no best-of, no night removed. Three of the 13 feeds duplicate another byte for byte and count once. Hence 10.
Nothing here is a recommendation to bet on anything.
Nothing to sign up for, nothing for sale. The log-in is ours.
These measure public weather models, not anyone's results.
42 nights is a small sample: the lead ranges from -0.16 to +0.22 °C. The middle of the field could reorder.
We trade these markets, and how we turn a forecast into a position is the part we keep. No prices, no returns, no rule here — only how well public models predict a public temperature. Anyone can reproduce it with the same free data.