PRISM

Property tax proposals stand or fall on counting. A surcharge on £2 million homes, a Council Tax revaluation, a replacement for Stamp Duty: pricing any of them means knowing how many homes sit at each value and where. No official source publishes that. Taxes fall on every home; price data exists only for the ones that sell. So we built PRISM, an estimate of England's full residential stock by current value and region.

England has around 26 million homes; about 830,000 a year sold across 2024 and 2025, roughly 3 in every 100, and nearly all hard price data comes from that self-selected slice. It is not a fair sample. A £350,000 semi in Leeds changes hands far more often than a £6 million house in Kensington, so counting expensive homes by counting expensive sales undercounts them, and the more expensive the home the worse the undercount.

The bias is measurable: compare each band's sales against the VOA's count of its homes, giving a turnover rate, sales per 100 homes per year.

The shape is a hump. Turnover peaks in bands D and E at about 3.7% a year, then falls: 3.1% in band G, 2.7% in band H. Band A is also low, at 2.3%; social housing, which almost never trades, is the audit's answer there. Both ends of the market are under-represented, so a correction must scale each cell's sales up by exactly how rarely its homes sell.

Every region has a hump, but not the same hump: the low-band shortfall deepens as you move south, and the top-end decline shallows. London is the outlier at both ends. Its band A homes sell at 0.6% a year, about a quarter of the national rate, while its band H homes trade more often than band H anywhere else in England.

The correction is a reweighting, done the way opinion polling does it. A pollster does not treat the thousand people who answered the phone as a miniature Britain; they reweight to known totals for age and region. We do the same for homes, anchored to the one source that covers every dwelling, sold or unsold: the Valuation Office Agency's count of homes in each Council Tax band, in every local authority. The table below is what that produces: pure counting, no fitted model in it. The rest of the page tests whether the counting deserves belief.

Two headlines. England has roughly 163,000 homes worth more than £2 million, almost two-thirds of them in London, and about 21,000 worth more than £5 million, 85% in London. At the other end, 1.35 million homes are worth under £100,000, seven in ten of them in the North East, North West and Yorkshire.

The reweighting works hardest at both ends: cheap homes sell rarely too (social housing, empty homes), so the sub-£100k rows are scaled up like the expensive ones. The difference is evidence: the cheap rows rest on tens of thousands of sales, the £10m-plus row on 280. That is why the top rows carry stated uncertainty: counted from 2024's sales alone the £10m-plus row reads 5,500, from 2025's alone 3,700. The published 4,700 averages two years that genuinely disagree, and the audit below measures that disagreement rather than hiding it.

2025 counts the top thinner than 2024 at every threshold, most at £5m+, the forestalling year having favoured the mass market; the published figures sit between the two as the average they are. 2023 would have read fatter still, by about 7% at £2m+ and 16% at £5m+. That is not evidence of more expensive homes in 2023: uprated by a regional index, 2023's band G and H sales come out about 3% high (see The stratification), and 3% on prices is enough to push some 20,000 homes over the £2m line, more than the whole excess. It is the reason 2023 tests the model rather than feeding the table.

PRISM sits underneath our published estimates. Wherever an analysis needs to know how many homes sit above a value, or what a rate schedule would raise, PRISM supplies the counts; the revenue figures in our High Value Council Tax Surcharge analysis are built on it.

A reweighting is only as good as its cells. If homes sold or stayed put for reasons the local authority and band grid does not capture, weighting by that grid would mislead.

So the second half of PRISM is an audit. We take the turnover differences above down to cell level and explain them with a regression, drawing on Census tenure data and VOA records of what the buildings are and when they were built. The regression never touches the table above (the weights are pure counting), but it is the reason to trust them.

The audit includes the strongest test a model like this can face: fit it on one year's sales, then ask it to explain a year it has never seen. Fitted on 2024 alone, the model explains 2025 almost as well as 2025 explains itself, and the reverse holds too. A third year, 2023, fails the same test on every pairing, which is why its sales inform the audit but are deliberately not counted in the table; the details are under Out of sample, below.

Taking the sales as they come understates the £2m+ stock by an eighth, because London and the south trade slowly. Most of that gap closes at the first rung: any weighting that knows a sale's region already scales prime stock up. The ladder then overshoots and comes back: region alone gives every London sale the same weight, though its expensive bands change hands faster than its cheap ones, so it over-counts the top; band corrects that, and the local authority corrects it a little more.

Two things follow. The England totals are not delicate: anything with band and region in it lands within a few per cent of PRISM. And band is what makes the answer stable from year to year, the unstratified and region-only counts moving 14% between 2024 and 2025 against 4 to 5% once band is in; the local authority adds no further stability, and its dividend is a count that holds at borough level, which is where tax proposals are costed.

The obvious explanation for low turnover at the top is price itself: Stamp Duty bills are largest there, and moving is costly. That turns out to be a small part of the story.

Across 2,163 local authority and band combinations, the strongest predictors are not prices but who holds the homes and what the homes are. The model measures twelve things about each cell: the ownership mix (social housing, private renting, the share owning outright with no mortgage, and residents' age); two markers of settled occupiers taken straight from Council Tax records, the share of households on pensioner Council Tax Support and the share of homes with a disabled band reduction; the buildings (flats, homes with four or more bedrooms, and the shares built before 1919 and in the 1920s); how fast the cell's stock is growing; and how much of it is annexes and park homes, which depress turnover even after their removal from the counts.

Social housing rarely sells on the open market; that is why band A turnover is low. Rented homes sell less often than owner-occupied ones. Mortgage-free owners, who skew old, are settled: their homes come back to market through deaths and downsizing, not job moves. Poverty pins people too: the poorest pensioner households, concentrated in the cheap bands of otherwise affluent districts, almost never move, and homes adapted for disability hold their households to the property. Buildings have habits of their own: Victorian and Edwardian stock churns, through conversions, landlords and first-time buyers, while interwar suburbs sit tight. And homes are sold by areas that build: where stock is growing, every new home has a first sale.

That is the case in one picture: social renting alone does four times the explanatory work of all six price terms combined, and price sits mid-table, level with pensioner poverty. Who holds the homes and what they are separates sellers from non-sellers; the next section measures what the little that price does looks like.

The most conspicuous absences are second homes and empty homes. Band H homes are more than four times as likely to be second homes as bands A to E, and in Westminster around 16% are second homes with another 3% empty: roughly one in five band H dwellings there is nobody's main residence. Striking as that is, neither status earned a seat: tested against the model, second-home and empty shares add nothing once ownership and residents' age are counted; the Tested and not adopted register below records the verdict.

The finding at the heart of PRISM. Measured naively, turnover falls sharply with price: a £5 million home sells a little over half as often as a £500,000 one. But account for what those homes are and who holds them, and much of the fall is composition rather than price. The model reproduces the whole band-by-band hump from composition plus smooth price terms.

What remains of price differs by region, and the model says so directly: it fits one price curve for London, one for the greater south (the South East, South West and East of England), one for the rest of England. Across the north and midlands, adjusted turnover declines only gently: a £5 million home there comes to market about five-sixths as often as a £500,000 one, once who holds it and what it is are held constant; strip out the covariates for pensioner poverty, adapted homes and Victorian-era churn and the decline deepens by half, composition masquerading as price (the specification table below shows it).

The greater south declines genuinely and more than twice as steeply, under every specification tried. London rises through the middle of the range and falls back below par by £5 million. Two-thirds of the sales in cells with a median above £2 million happen in London boroughs, so a single national curve would largely inherit London's answer and contradict everywhere else; PRISM does not publish one.

This matters for the table's top rows. PRISM never assumes a £10 million house sells at some ever-smaller rate; each cell's weight just counts what happened there. The finding here justifies that: what stops expensive homes trading is mostly measurable (ownership, residents' age, pensioner poverty, disability adaptations, building type and vintage, all from published data), so weighting by cells that hold those things roughly constant is a sound way to count unsold homes.

Two cautions. The exact height of these curves at £5 million is the least settled number on this page: fitted to single years it moves substantially, especially in London, where even the direction is not identified, and where letting the composition effects differ from the rest of England reverses it (see Stress-testing); the specification table below gives ranges. And every part of England faces the same Stamp Duty schedule, so comparing places cannot tell us what the tax does to moving; the note under the regression draws that line.

Price patterns alone explain about 37% of the variation in turnover between cells. The twelve composition covariates take that to 81%; letting the price curve differ by region group and the social and outright effects by price level reaches 84%. That is not a memorised fit: refitted with whole local authorities held out and predicted cold, the model still explains 83%.

The sharpest test is the hump itself: the model is never told which band a cell is in, yet it reproduces the whole national band-by-band curve, both low ends included, within 0.15 points.

Band averages can flatter a model; errors in opposite directions cancel. So here is every cell. The big cells sit close to the line; the scatter is widest among small cells, where a handful of sales swings the rate. The correlation is 0.85.

At the top end, borough by borough, the fit is rougher and misses both ways. The model expects prime central London to sell more rarely than it does: Camden, Kensington & Chelsea and Westminster all outsold their predictions across 2024–25, each by around 0.4 percentage points. Two of the caveats below are the likely causes: our neighbourhood-based tenure estimate probably overstates renting in prime London, and rumoured taxes on high-value homes may have pulled top-end sales forward. Birmingham and Richmond err the other way; Birmingham's band H homes sold at 1.6% against a predicted 2.4%.

The universe is sellable homes. The stock control is the VOA valuation list at 31 March 2025, minus the two dwelling types whose sales our transaction data can never contain: annexes, rarely marketable separately from the main house, and caravans, houseboats and mobile homes, whose sales bypass the Land Registry. That removes 175,700 dwellings, 0.7% of the list; 99% of them sit in bands A and B, none in G or H.

The exclusion matters more than its size suggests because it is concentrated: park homes are three-fifths of band A on Canvey Island, annexes more than a quarter of it in Waverley and Uttlesford. Left in the denominators, those dwellings would make affluent southern band A look almost saleproof, and the weighting would silently redistribute their stock across ordinary band A sale prices.

Cells and weights. Every sale is assigned to a local authority × Council Tax band cell; a cell's weight is its sellable stock divided by its sales over the window, matched plus the fractional assignments described next. Weighted counts sum to the VOA's sellable-stock totals by construction; the published table covers the 98.5% of stock in cells with at least one matched sale.

The roughly 7% of sales without a recorded band (15% in London) are not dropped: each is assigned a band probabilistically from its price and location, using the price distributions of matched sales in the same local authority: an unmatched £5 million Westminster sale might count as 0.9 of a band H observation and 0.1 of band G. These fractions enter the weighting and the turnover rates alike; only cell median prices use matched sales alone.

Values are Q2 2026 values. Every price is indexed forward to a common valuation date of Q2 2026; the sale year selects which transactions count, it does not date the values. The uprating uses Nationwide's regional house price index, a published standard series: each sale moves from its sale quarter to Q2 2026 by its region's index, so within a region the relative prices are exactly as transacted. Sales are sense-checked against a fixed plausible-value range for their region and band before they count (about 0.5% fail; details in the caveats below).

The counting window is 2024 and 2025, and it is chosen by test, not convenience. A candidate year enters the window only if its prices index reliably to the valuation date and the model fitted on it explains the years already counted. 2023, the peak-mortgage-rate year, fails both: uprated by the regional index its sales come out about 2% high against the same cells measured from 2025 sales, worst in band G at 3.4%, because a regional index cannot see prime stock underperforming its region. And its market traded differently enough that coefficients fitted on it explain 2024 and 2025 several times worse than those two years explain each other. So the dataset reaches back to January 2023, but 2023's sales test the model rather than feed the table (see Out of sample, below).

Turnover is a rate, not a share. Rates divide each year's sales by that year's own stock vintage and average across the 2024–25 window: every figure on this page is sales per 100 sellable homes per year. It is a transaction rate: 1.2% of a single year's sales, and 2.2% across the two-year window, are repeat sales of the same property, they belong in the numerator, so the rate is not "the share of homes that sold".

The unit of observation is a local authority and band cell, 2,163 in all. Each cell's 2024–25 sales count, matched plus fractional assignments, is modelled from its sellable-homes exposure, median price and twelve composition shares. There are no band dummies: any pattern across bands must come from the covariates.

Effects are not forced to be uniform. The price curve is fitted separately for London, the greater south and the rest of England, and the two strongest tenure effects, social renting and outright ownership, may differ between cheap cells (median under £300,000), mid-market cells and expensive cells (over £1 million).

The regression does not set the weights behind the table above; those are each cell's raw stock-to-sales ratio. It audits them: observable composition explains the turnover differences between cells, leaving price selection little to explain.

Four modelling notes. The price terms do no hidden work: replacing all of them with a free dummy per band reaches only 0.824 against the model's 0.840, so the three regional curves carry more structure than band dummies could.

Every covariate must earn its keep against held-out local authorities; the bar is stated under Stress-testing below, and the next section records what failed it, from the second-homes pair to the level of Council Tax itself.

One entanglement is accepted rather than hidden: outright ownership and residents 65+ are correlated at 0.93 across cells: read the pair jointly, as "settled elderly ownership", not either coefficient alone; dropping the age share costs 1.6 points of held-out fit. Stock growth is measured over the same two years as the sales it helps explain, so its coefficient bundles new-build first sales with any demand correlation; swapping in the built-since-2021 share costs little and changes nothing else.

And one honesty note on the 83%: the cross-validation folds were consulted during covariate selection, so it is marginally optimistic as a pure out-of-sample number; the cleaner test, fitting one year and validating on the other, is reported under Out of sample below, and the model passes it between 2024 and 2025.

What the regression is not: an estimate of Stamp Duty's effect on moving. Every cell faces the same tax schedule at a given price, so a cross-section cannot separate price from tax; measuring it means watching behaviour change when the tax changes, as the transaction-tax literature does (Besley et al.; Best and Kleven; Hilber and Lyytikäinen). The adjusted curves neither contradict nor support that evidence. The coefficients also describe places, not people: tenure and age come from neighbourhood data, so read them as area-level associations.

The covariate table above lists what holds a seat; this is the record of everything that was refused one, and why. The scoreboard: 12 covariates hold seats from 42 candidates given a hearing, and the 30 rejections are grouped below.

Every fit statistic is an invitation to ask what would break it. This section collects every way we attack the regression; the two that need tables of their own (the specification battery and the year-by-year tests) follow as their own sections.

The drop-one column measures fit, not what each covariate does to the adjusted price curves, which are the audit's key exhibit at the top end. So here is each regional curve's £5m point under alternative specifications, on two bases: relative to £500k as elsewhere on the page, and relative to £1.5m, because every curve falls from £1.5m to £5m and it is that stretch which decides whether transaction counts understate the standing stock at those values.

Five readings. The stripped-covariate row is the important one: remove the Council Taxbase pair and the pre-1919 share and the Rest appears to decline to 76 while London jumps to 95; with them measured, the Rest's decline is gentler (83.8) and London sits below par. The greater south declines in every variant, and is the least settled in level (35 to 54 in every other variant, 59 on the well-measured cells). The rest of England is a gentle decline in every other variant (74 to 90, 92 refitted only on cells with at least 50 sales). London sits below par at £5m in every other variant (72 to 82).

The £1.5m base tells one story outside London: past £1.5m the rest of England falls gently (94) and the greater south hardest (69), so transaction counts understate the standing stock above that point. London’s 77 does not survive the tests under Stress-testing: refitted on the top three bands alone its curve above £1.5m is about flat, and with London-specific composition effects it rises. And every £5m point is read beyond the last cell median (£3.2m in the rest of England, £3.4m in the greater south, £4.1m in London); the £2m points, which sit inside the data (87, 72 and 99), tell the same story. The deepest uncertainty is between years, not between specifications, and the next section measures it directly. Treat the £5m points as indicative: the greater south's decline replicates across three years of sales, the Rest's gentle decline across the two counted years; in London even the sign does not.

Every fit statistic above is measured on the sales the model was fitted to. The stronger question is whether the structure survives contact with sales it has never seen. Two tests answer it: fitting the full specification on each calendar year alone, and then asking each year's coefficients to explain the other years' cells.

The regional pattern replicates. In every year the greater south declines from £500k to £5m (52 to 72 across three very different market years), and the rest of England declines gently in both counted years (90 and 82) while sitting near par in 2023 (97). London does not replicate: 72 to 104 across the same years, so its top-end direction remains open. 2023 also stands apart in level, its top end trading relatively harder everywhere, and in fit: the composition covariates explain less of the peak-mortgage-rate year, 0.77 against 0.82.

2024 and 2025 validate each other: coefficients fitted on either explain the other to within about 0.03 of its own fit, which is as close to a pass as an out-of-sample test of this kind gives. Every pairing involving 2023 loses 0.11 to 0.17, three to five times as much, in both directions; 2023's market is structurally different, not merely noisier. That asymmetry, together with the indexation drift under The stratification, is the case for the counting window: the published table averages the two years that agree, and states the third as evidence.

The regional humps at the top of the page are not one shape with nine labels, and one set of coefficients cannot describe them. Refitted alone on a plain uniform specification, the regions disagree sharply: seven of the nine put the adjusted £5m relative between 31 and 65, the North East sits above parity, and London puts it far above, at more than twice the £500k rate. That disagreement is why the price curve varies by region group and the strongest tenure effects by price level; the question for the audit is how much of each region the single fitted model explains.

The model explains most of the cell-to-cell variation everywhere: 0.85 to 0.93 across the north and midlands, 0.74 to 0.81 across the greater south, 0.76 in London. The south is weakest; that is where the remaining heterogeneity lives. The interactions cover the price curve and the two biggest tenure effects; letting more vary would continue down the same road at the cost of many more parameters.

The tripartite grouping follows how regional house-price indices have diverged since the 1980s. Fuller interaction designs, effects varying by band group within each regional group, were tested during selection; the price-group form generalised to held-out authorities, the region × tenure form did not; the record is in the model repository.

Any cross-sectional regression invites the question of what its covariates really measure. Three mechanisms deserve naming:

The coefficient table above reports the composition covariates and their interactions; this is every parameter of the fitted model, with its standard error clustered by local authority, written out by the fitter at each rebuild. The same figures are available as JSON.

We build PRISM from public data and our own transaction records:

• Sales and valuations: 2.8 million Land Registry transactions from January 2023 to June 2026, matched to Council Tax bands; the model counts calendar 2024 and 2025 and reports the annual average, with 2023 held back as out-of-sample evidence. Every price is indexed to Q2 2026 using Nationwide's regional house price index: a 2025 sale counts at what the home is worth now, not what it fetched then.
• Stock counts: VOA Council Tax stock of properties (CTSOP), band counts down to neighbourhood level, 31 March 2025.
• Dwelling type and build period: the same VOA release, broken down by property type, bedrooms and build period for every band and local authority; also the source of the annexe and park-home counts behind the sellable-homes universe.
• Council Taxbase: the local-authority-level workbook, band by band for every English authority (October 2025 snapshot, January 2026 revision). Source of two model covariates, the pensioner Council Tax Support share and the disabled band-reduction share; of the second-home and empty-home shares quoted in the lay half (descriptive, not covariates); and of the Council Tax exemption shares that downstream revenue estimates set aside.
• Tenure and age of residents: Census 2021 owner-occupation (outright versus mortgaged), renting shares and age structure by neighbourhood, combined with the band counts.
• Geography: ONS lookups linking neighbourhoods to local authorities and constituencies.

• Tenure is estimated, not observed. No dataset records tenure by Council Tax band directly. We estimate it from neighbourhood-level data: reliable in most places, but it probably overstates renting among the most expensive central London homes, the leading suspect for the underprediction of prime London band H sales.
• Years disagree at the top, and the disagreement is measured. Fitted year by year (see Out of sample), the greater south's decline holds across three very different market years and the rest of England's gentle decline across both counted years, while London is unstable, 72 to 104. Each year carries its own distortion: April 2025's Stamp Duty threshold change moved completions visibly, 138,000 in March against 41,000 in April, inflating mass-market volumes and with them the £500k baseline; 2023's mortgage-rate peak suppressed the leveraged mass market while cash-heavy prime kept trading. The model averages 2024 and 2025 so neither sets the answer alone. The top-end gradient's direction is identified outside London; in London it is not, and the page says so wherever it appears.
• Sales are sense-checked before they count. Every sale must fall inside a plausible value range for its region and Council Tax band, or we drop it as a data or address-matching error. About 0.5% fail, in two very different ways. A handful of impossible high-value sales in low bands, a £22.8m band B in Portsmouth, an £18.0m band A in East Suffolk: tiny in number, but the reweighting multiplied each into dozens of phantom homes at the very top. And a larger group of genuine transactions that are not sales of a whole home: lease extensions, shared-ownership tranches and staircasing, recorded at a fraction of the dwelling's value. Thresholds are fixed constants, set once from the sales distribution rather than recomputed from the data being filtered, so outliers cannot inflate the range meant to catch them. One consequence: no London band admits a sale under £100,000, so the table shows no London homes below that. That is the threshold speaking, not a finding.
• Coverage. PRISM covers England only; Welsh Council Tax bands date from a 2003 revaluation and would need separate treatment. Around 7% of sales cannot be matched to a band, 15% in London; via the probabilistic assignment under The stratification, that share of the turnover numerator is imputed rather than observed, least certain in prime London. And the sellable-homes universe excludes annexes and park homes: real dwellings whose values PRISM cannot observe, so they appear in the note beneath the output table rather than its rows.

Poisson regression across 2,163 local authority × band cells covering 98% of England's sellable stock; fit shares are deviance explained; the out-of-sample figure holds out whole local authorities in five-fold cross-validation.

We are grateful to John Muellbauer at Nuffield College, Oxford for detailed comments on the model and wider discussions through the past year. The probabilistic treatment of unmatched sales is his suggestion, as are the tripartite regional grouping and the interaction programme built on it. The £1.5m-rebased column, the joint dead-stock constraint and the like-for-like drop-one comparison answer questions of his, as do the weighting ladder under Why believe this table, the year-by-year count of the top of the table, the test of observed stand-ins for the ecological block and the full listing of the regression; so do the London-specific composition effects, the −1 floor and the bands F to H refit under Stress-testing. His persistent scepticism of what ecological tenure coefficients really measure shaped the caveats. All errors are ours.