This page does not forecast anything. It shows what 1960–2025 of measured UK weather did to the two perils temperature actually reaches, what that evidence changed in the model, and — the harder half — what it deliberately did not.
Warmer summers dry clay soils, clay soils shrink, and buildings on them subside. Milder winters freeze fewer pipes, and frozen pipes are what turns a cold snap into an escape-of-water claim. One peril gets worse as the country warms and the other gets better, so "climate change raises premiums" is not a statement this model can make without saying which peril it means.
Subsidence is 12.1% of modelled claim cost. Its frequency now carries a drought climatology built from the same daily grids as the chart below — the model's subsidence map used to be geology alone.
Escape of water is 25.8% of claim cost, and 31% of it rides on frost days. Frost has fallen by a fifth since the 1961–1990 normal — and the model is blind to that by construction. Why, and whether that is defensible, is below.
The index is the peak of each year's running soil water deficit — rainfall minus evaporative demand, accumulated from 1 January and reset every year. It is computed per postcode district from HadUK-Grid 1 km daily grids and weighted here by households, so the national line is what an average policy experienced rather than what an average square kilometre did.
2,736 postcode districts · millimetres · marked years are the surges UK subsidence is canonically dated to
annual peak deficit least-squares trend canonical subsidence year
The trend is upward — +3.6% of the mean per decade (p = 0.036), against a 1991–2020 average of 260 mm — but the trend is not what earns the index its place in the model. This does: 5 of the 6 canonical UK subsidence surges land in the index's own top ten years (1976, 1995, 2003, 2018, 2022), and it was never fitted to them. The miss is 2006.
On that evidence the subsidence frequency surface stopped being geology alone. Each district's frequency is now multiplied by a relativity that is 43.5% flat and 56.5% drought climatology, the 56.5% being the ABI's own attribution arithmetic rather than a tuned number. The national loss cost cannot move — the calibration re-pins it — so the change is purely about who carries it: 455 of 2,736 districts changed rating group, 1 of them by two, and the headline premium moved by less than a penny, to £169.73. How the subsidence surface is built →
The same extraction gives air-frost days, the index behind the freeze-sensitive part of escape of water. The decline here is the clearest temperature signal in the whole dataset.
2,736 postcode districts · days
air-frost days least-squares trend
Frost days are falling at 7.5% of the mean per decade (p = 0.0002): 49.0 days a year across 1961–1990 against 39.0 across 1991–2020, a fall of 20.5%. And the model cannot feel any of it.
That is arithmetic, not an oversight. Each district's freeze exposure enters as its frost days divided by the national household-weighted mean, so a decline that is roughly proportional across the country cancels exactly. Only two things could move a premium: the shape of the frost map, and the share of escape of water assigned to freeze at all.
The shape was tested against the obvious objection — that a climatology measured over 1991–2020 is out of date — and it survived. Re-aiming the window at the most recent twenty years moves the relativity map less than two halves of the current window move it against each other. A shorter, more recent window is a noisier estimate as well as a more current one, and here the noise is larger than the signal: the difference between the 1961–1990 frost map and today's is no bigger than the difference between two arbitrary decades inside the current one. Re-aiming would buy a worse estimate of the same map, so the map stays.
That leaves the share — and rather than argue about it, it was priced. Re-running the full model across a 0.20–0.40 range, nearly a doubling of the dial, moves the national premium by at most 0.17 pence. The expected loss for escape of water does not move at all: it is identical to the last bit of floating-point precision in every run, because the relativity is normalised and the calibration re-pins the level. Re-aiming the climatology window costs 0.02 pence, which is the measurement above confirmed by simulation rather than merely asserted.
What the dial does do is decide who pays. At the low end 353 districts change rating group, and the largest single-district move is about £15 a year — Highland districts such as Inverness-shire falling, the Isles of Scilly rising, which is exactly the shape you would expect from turning frost sensitivity down. So this is a distributional question wearing a pricing question's clothes, and there is no evidence with which to answer it: no UK publication breaks escape-of-water claims down by freeze by era. The share stays at its anchored value and the gap is recorded as a limitation rather than filled with a judgement. The caveats list →
Temperature is not the only way weather reaches a premium, and the model has a direct measurement of what separates a bad year from a normal one. It is not what most people assume. A catastrophic year — the worst 1% the simulation produces — carries 1.40× the claim count but 2.60× the claim value of a typical year. Bad years are made of dearer claims, not many more of them; the count saturates while the value keeps climbing.
Of that 2.60×, about 1.22× is the mix of claims shifting (storm falls from 74.2% to 66.5% of claims while flood rises from 9.7% to 14.6%) and 2.13× is severity rising within perils. Storm is nearly flat at 1.26× count and 1.07× value; flood is the engine, at 2.12× count and 2.90× value. The year analysis →
It is not a forecast. Every number here is measured over 1960–2025. No temperature projection is applied anywhere in the model, and the two trends above are descriptions of what happened, not extrapolations of what will.
The climate scenario on the map is flood only. The climate repricing uses the Environment Agency's own future flood extents. It contains no drought term and no freeze term, so it cannot show subsidence worsening or escape of water easing. The uplift it reports is a flood uplift, and reading it as a total climate effect would overstate one direction and ignore the other.
The drought index is a relativity, not a soil-physics quantity. Its evaporation term uses the Hargreaves–Samani formula, which runs about a third high in a maritime climate. That was measured rather than assumed: re-running the integrals at 85% and 70% of the evaporation moves the level but leaves the map almost exactly where it was (rank correlation ≥ +0.998), and the level is precisely what the calibration re-pins. Only the map is used.
One number in the freeze half has no anchor at all — the share of escape of water attributed to freeze. It is carried at its published value, and the section above says why it was not adjusted.
Both series on this page come from HadUK-Grid 1 km daily observations (Met Office, via the CEDA archive, Open Government Licence), integrated over every one of 2,736 postcode districts and weighted by census households. The daily grids are roughly 174 GB and are never stored whole: the extraction streams them a year at a time and keeps only the annual reductions.
Every figure on this page is injected from the committed series at build time. None of them is typed into the page, because a number typed into a page is a number that goes stale. How to rebuild the whole thing →