Investor Properties

Valuation accuracy

How accurate are our valuations?

Every month we check our automated valuations against what homes actually sold for, on recent sales our valuation model had never seen.

How valuations feed into each deal's verdicts and score, and every change to those rules: how a deal is scored.

In short

  • 54% of our estimates were within 10% of the sale price in our latest test (13 November 2026), on sales our model had never seen.
  • Every estimate comes with a likely range; 81% of sale prices fell inside it.
  • Confidence matters: Good-confidence estimates land much closer to the sale price than Low ones, so every deal shows its confidence.
  • Use them to sort listings down to the few worth a closer look, not to decide what to pay: check the comparable sales and get a survey.

Current model

  • Valuation model v2 (gbm-2026-11-05-v2), learned from Land Registry sales and energy certificates up to September 2026.
  • Since 13 Nov 2026: a correction for studios and flats valued from the street (below); a typical floor area from energy certificates on the same street for listings with none of their own; and, where an independent valuation is lower than ours, deal figures use the lower of the two.
  • Which figures include those changes: the listing counts, the mix the headline is weighted to and the second opinion check are live, so they do. The test results (the headline, the bands and the results by confidence) come from the test run on 13 November 2026, which values each past sale the way we now value listings: with the studio and flat correction and street floor areas. The 6 Nov result, on the model before those changes, is in the log of changes below.

Latest result

54.1%±0.1 points

of our estimates were within 10% of the sale price, across 592,976 sales in England and Wales.

Margin of error: 95% confidence interval 53.9% to 54.2%. Tested on 13 November 2026, on standard home sales from 30 September 2025 to 30 September 2026.

Why estimates like these are still useful

Most adverts don’t give the address, so about half our estimates land within 10% of the eventual price. That’s not precise enough to decide what to pay, and we don’t ask you to. What it is good for is sorting hundreds of new listings down to the few worth a closer look, and being clear about how far to trust each figure.

  • Every estimate says how sure it is. Good-confidence estimates were within 10% for 64% of sales; Low-confidence ones for 28%. Each comes with a likely range, and 81% of sale prices fell inside it (we aim for 80%), so you can see when a figure is a rough guide.
  • The checks that depend on the value are cautious. A deal is only rated strong when we’re confident in the value, and deals where a second valuation disagrees with ours, or where the price looks too far below our value to be true, are flagged and never featured or sent as same-day alerts. Land, plots and commercial property aren’t valued automatically: you can still work out their figures with your own numbers.
  • Your own figures come first. On every deal page you can replace our value, the price, the cost of works and the rent with your own, and every figure is worked out again.
  • Some things only a visit can tell you. No automated estimate can see inside a home: its condition, its layout, or a problem with the plot. That’s what viewings, surveys and your own research are for.

Studios and flats valued from the street

Studios. We send each listing's bedrooms with it, and since 13 Nov 2026 a correction recognises studios (no separate bedroom) and lowers their values: ×0.771 when we only know the street and ×0.900 with a floor area, with ranges about 30% and 19% wider. Values made with the address aren't corrected.

Before the 13 Nov correction (testing on 1,544 studio sales, 6 Nov): values made from the street alone came out a median 28% above the sale price, with 20% within 10% of it; with a floor area, 11% above and 36% within 10%; with the address, 4% above and 47% within 10%.

After the correction (on the half of the test year it wasn't fitted to): values from the street alone came out a median 2% below the sale price, with 31% within 10% of it; with a floor area, 2% below and 42% within 10%; and the ranges again catch about 8 in 10 sale prices. Studio values are still rough, so every studio's value is still shown at Low confidence, and we say only whether its price is below, within or above our value range, never by how much.

Other flats valued from the street alone were not biased before the correction (a median 0.4% above the sale price in the 6 Nov test; the 13 Nov correction moves them by only ×0.997, with slightly wider ranges), but they're less precise: 40% within 10% of the sale price, against 55% when we know the address (6 Nov test, before the correction). Many flats in the deal feed are listed far below our value because of things we can't see, such as a short lease, condition or a repossession, so check those before relying on the figures.

Full results and method

Latest test in detail

The headline 54.1% is the share of 592,976 sales where our estimate was within 10% of the price, weighted to what today’s listings tell us: 17% give an address we can use, 68% a street and floor area, and 15% only a street.

The test and this weighting cover standard homes only: houses, bungalows, flats and maisonettes, sold at market value. Land, plots, development sites, garages, parking, commercial and mixed-use listings aren’t valued automatically (142 of the listings we hold today), so they’re left out of both.

81% of sale prices fell inside our range (target 80%), and 80% when the address was known. Every valuation comes with an 80% range: 8 in 10 homes should sell inside it.

Within 5%
29.9%
Within 10% (the headline)
54.1%
Within 15%
70.4%
Typical error (median)*
9.3%
Over or under, on balance*
+0.0%

The within figures are weighted to today’s listings, like the headline. *A median can’t be weighted that way, so these two are for every sale valued from its street, with the floor area where known: 53.3% of those were within 10%.

Valuation model v2, Nov 2026 · Run on 13 November 2026 · sales from 30 September 2025 to 30 September 2026 · 592,980 sales tested, 4 without enough nearby sales to value.

How far our estimate was from the sale price
  • Within 5%29.9%
  • 5–10%24.2%
  • 10–15%16.3%
  • Over 15%29.6%

Weighted to today’s listings, like the headline: the first two bars add up to the 54.1% within 10%, and all four to 100%. Bands are the difference either way from the price.

By what we know about the home

Each test sale is valued three ways, as we would value a listing that gave its address, one that gave its street and floor area, and one that gave only its street.

By what the listing tells us

SalesTypical errorWithin 5%Within 10%Within 15%
With the address17% of today’s listings592,9767.7%35%61%±0.177%
Street and floor area68% of today’s listings (tested where the floor area is known)549,7238.8%31%55%±0.172%
Street only15% of today’s listings293,94513.2%21%40%±0.255%

We can value 17% of the 2,272 listings we hold today with their address: the advert gives a house number, or we have matched the advert to its address (only to value it; we never show an address the agent didn’t publish).

Second opinion check

On shortlisted deals and listings that match our members’ searches, we also get an independent automated valuation from our licensed data provider and compare it with ours. This measures how often two models agree, not accuracy: neither is the sale price. Where they’re more than 10% apart, the deal page shows an amber flag and the deal can’t be rated strong.

Our valuation against the independent one, current listings

ListingsMedian differenceMedian gap, either wayWithin 5%Within 10%Within 15%
All listings268−8.2%12.4%22%40%60%
Unmodernised homesour value adjusted for condition72−10.1%14.2%17%33%51%
Other homes196−8.0%11.7%23%43%63%

Valuations from the last 60 days, latest 19 November 2026. Median difference is the independent value against ours (minus: it is lower). We compare with our value as the site shows it, which for unmodernised homes is adjusted for condition; the independent model is told the condition too.

What the gap means (13 Nov 2026)

On 13 Nov the independent valuation was a median 9.3% below ours, across 152 listings. Most of that is which listings get checked: we only ask for a second opinion where our value is already well above the asking price, and the further below our value a listing looked, the bigger the gap (13% where it was priced 20% or more below ours, and slightly the other way where it was priced above ours). So on the deals we show, our value is probably about 10% too high at the median. Flats (13% lower) and detached homes (19% lower) account for most of it; semis and terraces were within about 1%. The gap was smaller at Good confidence (2%) than at Low (13%). Tested against what homes actually sold for, our estimates show little bias overall, so treat a large discount to our value as a reason to look closer, not as a guaranteed margin.

Why the counts on this page differ

  • Sales (592,976 and so on) are past Land Registry sales in the test run on 13 November 2026, not listings. 592,980 were tested; 4 had too few nearby sales to value. Street and floor area is tested only on sales whose floor area is known, so it has fewer.
  • 2,272 listings (live, updated hourly) are the standard homes for sale we hold today that our model values, including the 17% we can value with their address: the headline is weighted to them, so it moves a little as listings come and go. We hold 2,431 listings in all: these 2,272, 142 land, plot and commercial listings (not valued automatically), 14 park homes, retirement or shared-ownership homes (not analysed as standard homes), and 3 valued from comparable sales because our model couldn’t value them.
  • 268 listings in the second opinion check: only shortlisted deals and listings matching our members’ searches get an independent valuation, and only those from the last 60 days count (live; latest 19 November 2026).
  • 2,633 listings underwritten: every listing we’ve underwritten so far, including those since sold or withdrawn, so it’s more than we hold today (live, a running total).
  • Counts in the model changes below (such as 526 or 846 listings) are the listings we held on the date of that entry.

What this does and doesn’t test

This tests our current-value estimate: what a home is worth as it stands, against what it actually sold for. It isn’t a test of our after-works values (what a home might be worth once refurbished), or of any particular deal’s figures. Every valuation on the site is an automated estimate, not a RICS valuation or survey, and nothing here is financial or investment advice.

By confidence

Every valuation comes with a confidence level, from how wide its 80% range is. Good: typically within about ±7% of the sale price. Fair: about ±11%. Low: wider than that (about ±18%). Here’s how each level did.

By confidence: street, with the floor area where known

SalesTypical errorWithin 5%Within 10%Within 15%
Good307,2087.3%36%64%±0.281%
Fair208,02910.8%25%47%±0.264%
Low77,73918.8%14%28%±0.341%

By confidence: with the address

SalesTypical errorWithin 5%Within 10%Within 15%
Good418,7346.5%40%69%±0.185%
Fair119,44410.8%25%47%±0.364%
Low54,79818.7%15%29%±0.442%

By region

SalesTypical errorWithin 5%Within 10%Within 15%
London59,3229.2%29%54%±0.470%
South East100,3768.4%32%57%±0.373%
South West67,5769.4%29%53%±0.469%
East of England67,3918.3%32%58%±0.474%
East Midlands52,0439.0%30%55%±0.471%
West Midlands53,7909.3%29%54%±0.471%
Yorkshire and The Humber57,30510.0%27%51%±0.467%
North West76,50010.0%27%50%±0.467%
North East26,96210.8%26%47%±0.664%
Wales31,71110.8%26%47%±0.563%

By property type

SalesTypical errorWithin 5%Within 10%Within 15%
Detached148,3339.6%29%52%±0.368%
Semi-detached181,6128.6%31%57%±0.273%
Terraced170,0618.9%30%55%±0.271%
Flats92,97010.9%25%47%±0.363%

Value-add stock and homes after works

Most of the deals we flag are unmodernised, reduced, repossessed, short-lease or auction sales, so we also test those homes on their own, whatever the results show. Land Registry data has no condition field, no lease length and no auction marker, and repossessions are mixed in with other sales we don't value, so we can't test repossessions, short leases or auction lots directly. Instead we use the closest proxies the data supports: homes whose energy certificate was rated F or G when they sold (usually unmodernised), and homes resold within a year of their previous sale (often after works).

Value-add proxies, valued from the street (and floor area where known)

SalesTypical errorWithin 5%Within 10%Within 15%
EPC F or Gunmodernised proxy6,83716.0%17%34%±1.148%
EPC A to Efor comparison542,9468.8%31%56%±0.173%
Resold within a yearof the previous sale2,51610.6%25%48%±1.965%

Homes after works. To test the value after a refurbishment, we take homes sold twice where a new energy certificate between the two sales rated them at least two bands better, and compare our estimate with the second sale (the price after works). Homes whose rating didn't change are shown for comparison. We can only pair sales whose certificates we could match to the home, so this is a subset of resales.

Resales after an EPC improvement

SalesTypical errorWithin 5%Within 10%Within 15%
EPC 2+ bands betterfrom the street5,9879.8%27%51%±1.370%
EPC 2+ bands betterwith the address5,9877.8%34%61%±1.279%
EPC unchangedfrom the street85,4138.5%31%57%±0.375%

Listings we underwrote, against what they later sold for

Coming soon. When we underwrite a listing we keep what we said: our estimate, the independent valuation where we had one, and the asking price. As those homes complete and appear in Land Registry data (usually a few months after the sale), we’ll compare all three with the price they actually sold for, and publish it here automatically once there are at least 100 matched sales; so far there are 0.

2,633 listings underwritten since 5 November 2026. Land Registry’s published data currently runs to sales on 30 September 2026. We match each sale to the listing by postcode and house number, never by price, and only sales after we first saw the listing count. Matches where the advert gave only the street are counted but not used here (0 so far).

How we test

  1. Our valuation model is trained on standard home sales from HM Land Registry since 2015, matched to the energy certificate (EPC) each home had before it sold. Everything it learns from a sale comes from before that sale: nearby prices in the months before, the local price trend, the postcode’s and street’s earlier sales, and the home’s own earlier sale and certificate.
  2. The split is by time. The model learns from older sales, its ranges and confidence levels are set on the six months after those, and it’s tested on the 12 months after that: sales it has never seen. Nothing from the test months is used to build or tune it.
  3. Each month we test every one of those sales across England and Wales: 592,980 in the latest run. (Until November 2026 we tested a random sample of 20,000, spread by region and property type in proportion to how many homes sold. Both measure the same thing; testing all of them makes the figures more precise.) New builds, and sales that aren’t at full market value, are left out, as they are when we value listings.
  4. Each home is valued as at the month before its sale, the way we value a listing: with only its street, with its street and floor area, and with its address. The sale itself is never used, and the asking price is never an input to any valuation, on listings or here.
  5. We compare each estimate with the actual sale price, and check how often the price fell inside our 80% range.
  6. The model is retrained every month on the latest data. A new version only replaces the current one if it does at least as well on sales neither has seen.
  7. Each within-10% figure comes with its margin of error (a 95% confidence interval, shown as ± points): the range the true figure is very likely in, given how many sales were tested. The headline’s margin is the cautious upper bound for a figure weighted across the three ways of valuing.

Sold prices: HM Land Registry Price Paid Data, Crown copyright, Open Government Licence. Energy certificates: the EPC register (MHCLG), Open Government Licence. More on where our data comes from.

Model changes

Each change we keep, with its effect on the same set of test sales.

  1. 5 November 2026The test now values each sale exactly as we value listings, including floor areas from energy certificates (EPC). A fairer test, not a model change: typical error 12.3% → 11.9%, within 10% 43% → 44%. It showed estimates were 3.5% low on balance.
  2. 5 November 2026When a home’s size-based value and its postcode’s earlier sales disagree by more than 30%, we now use a value between them instead of the lower one. The main cause of the low bias: typical error 11.9% → 10.6%, within 10% 44% → 48%, bias −3.5% → −1.3%.
  3. 5 November 2026Small calibration factors by region and property type (at most ±4%), fitted on a separate, earlier sample, for the bias that was left. Typical error 10.6% → 10.5%, bias −1.3% → +0.2%.
  4. 5 November 2026We switched to our valuation model: a machine-learning model trained on 7.9 million Land Registry sales from 2015 and the EPC register, tested on later sales it never saw. It never uses the asking price, and gives each value an 80% range. Within 10%: street, with the floor area where known, 48% → 54%; street only 42% → 45%; with the address 50% → 61%. Typical error with the address 10.0% → 7.7%. 80% of sale prices fell inside our range (target 80%).
  5. 6 November 2026We now find more listings’ addresses, only to value them (never shown): from the energy certificate register when exactly one home on the street fits the advert (right 98.6% of the time when tested on 13,615 past sales), and from our licensed listings provider for Rightmove listings. Listings valued with their address: 20.2% → 21.9% of the 526 we held.
  6. 6 November 2026Condition adjustment for unmodernised homes, calibrated on 17,153 repeat sales where the energy certificate showed a renovation between sales: about −10% on average, varying by property type and EPC band (at most −20%). It applies only to listings we class as unmodernised, from the advert (or an F or G energy rating), and never to the value after works. Adjusted 58 of the 545 listings we held. The back-test results above are unchanged: they measure the model itself on past sales, which aren’t classed by condition. Checked against an independent automated valuation on 30 listings picked at random (not by our own value), with condition sent: for 15 unmodernised homes, the independent valuation was a median 17.9% below our value before the adjustment and 12.6% below after, against 5.7% below for 15 other homes.
  7. 6 November 2026We now read a home’s floor area from the advert when the agent states it for the home as it stands (not a plot, an outbuilding or a planning permission). It matched our listings provider’s floor area within 10% on all 28 listings that had both. Listings valued with only the street: 3 of 544 more now have a floor area. Within 10% for the homes we list: 54.3% → 54.4%.
  8. 13 November 2026A correction for studios and flats valued from the street, because studios came out a median 28% above the sale price from the street alone (11% with a floor area). Fitted on the first half of the test year and checked on the second: studios ×0.771 from the street alone and ×0.900 with a floor area, with their ranges widened (×1.30 and ×1.19); other flats ×0.997, ranges ×1.03 to ×1.05. Values made with the address aren’t changed. Each listing’s bedrooms are now sent with it, so the correction can tell a studio from a one-bedroom flat; studios stay at Low confidence. Before it, studios from the street alone were 20% within 10% of the sale price (36% with a floor area); after it, on the half of the year it wasn’t fitted to, 31% (42%). Re-valued all 846 flats and studios we held: studios a median 10.0% lower, with ranges 18% wider; other flats 0.3% lower. The test results above include this correction (re-run on 13 November).
  9. 13 November 2026For listings with no floor area and no address, we now estimate a typical floor area from the energy certificates of similar homes on the same street (the same kind of home, with rooms that fit its bedrooms), used only when at least two certificates agree. It values the home and is never shown as the home’s floor area, and its value range is widened to allow for it (×1.04, or ×1.16 when the certificates agree less closely). From 13 Nov 2026 it also sizes the works estimate for a home with no floor area of its own (high and medium agreement only), instead of a typical size for the bedrooms. Tested on 125,690 test-year sales: within 10% rose 2.3 to 2.5 points against valuing from the street alone. Of the 554 listings we valued from the street alone, 196 are now valued with a street floor area.
  10. 13 November 2026

    13 Nov: headline 55.2% → 54.0% within 10%

    The drop comes mostly from a fairer test, not a worse model. Homes whose advert gives no floor area are now tested the way the site actually values them, from the street’s typical floor area rather than their true one. And “street only” homes are now the harder cases left over.

    • Valued with the full address: 61.0% (unchanged)
    • Street plus floor area: 56.1% → 55.4%
    • Street only: 44.7% → 39.9%
    • Sale prices inside our 80% range: 80.6% → 80.7%

    Using the street’s typical floor area beats using the street alone for every type of home: 52.0% vs 49.5% within 10% overall, and studios improve most (41.9% vs 31.6%).

Earlier results

RunMethodSales periodSalesTypical errorWithin 10%
13 November 2026Valuation model v230 September 2025 to 30 September 2026592,9769.3%54.1%±0.1
6 November 2026Valuation model v230 September 2025 to 30 September 2026593,9179.1%55.3%±0.1
6 November 2026Valuation model v230 September 2025 to 30 September 202620,0009.1%55.1%±0.7
5 November 2026Valuation model v230 September 2025 to 30 September 202620,0009.1%55.1%±0.7
5 November 2026Comparable sales30 September 2025 to 30 September 202619,98510.5%48.4%±0.7all sales

Within 10% is the headline figure for each run: our valuation model’s results weighted to what today’s listings tell us (comparables runs, before the model: all sales). Runs before November 2026 tested a random sample of 20,000 sales, spread by region and property type; later runs test every held-out sale (592,980 in the latest). Both estimate the same figures, so they can be compared; the ± margin (95%) shows how precise each run is: about ±0.7 points for 20,000 sales and ±0.13 for 594,000.

Estimates only. Every valuation on the site is an automated estimate, not a RICS valuation, a survey or a lender’s valuation, and nothing here is financial or investment advice.

Contains HM Land Registry data © Crown copyright and database right, and energy certificate data, licensed under the Open Government Licence v3.0.