Accuracy
Quarterly bulk test results against the hardest benchmark: actual Land Registry sale prices. No cherry-picking. No simulated data. These are real predictions vs real outcomes.
Fitch Band B
All figures against actual Land Registry sale prices. Gadsden Meridian v6.13 · H2-2025 bulk test of 295,026 transactions.
Benchmark basis (from 15 June 2026, Meridian 6.12.1). Headline accuracy is reported on standard market sales — Land Registry Category A: PE10 67.4% (on 266,525 standard market sales). On the full set including non-standard sales — repossessions, buy-to-lets and corporate transfers (Category B) — it is 65.7%. Category B sales transact under conditions the model is not built to price; the blended figure is retained in the technical document.
Note on bias. From Meridian v6.12, bias is the median signed percentage error — the robust measure adopted alongside FSD per Option A (EAA ESSVM convention). v6.11 figures were the mean signed percentage error. This is a measurement-definition change: it can move the number, and occasionally flip its sign, without any change to the model’s predictions. Same valuations; bias is now reported on the robust basis the standard uses.
Most AVMs are benchmarked against surveyors’ valuations. We test against actual sold prices — a harder benchmark.
Industry thresholds were originally defined against surveyor valuations — a softer test. Our 67.4% PE10 on standard market sales against actual sale prices exceeds the Kirchmeyer minimum.
How we benchmark
We measure against the actual Land Registry sale price — the amount the buyer paid and the seller accepted. A tighter, taller curve means more valuations land closer to the real price.
* Hometrack benchmarks against surveyor valuations, not actual sale prices — a less demanding test.
Industry thresholds
PE10 — the percentage of valuations within ±10% of the actual sale price — is the standard accuracy metric for AVMs. Academic and industry research defines several recognition thresholds.
All figures benchmarked against Land Registry sale prices. Industry thresholds were originally defined against surveyor valuations — a softer test.
Reading the metrics
PE10 (Percentage within 10%)
The proportion of valuations where the model's estimate fell within 10% of the actual sale price. This is the industry-standard accuracy metric for automated valuation models. A PE10 above 80% is considered strong performance for a national AVM.
MdAPE (Median Absolute Percentage Error)
The median percentage by which the model's estimate differs from the actual sale price. Using the median rather than the mean prevents outliers from distorting the figure. Lower is better.
Test count
The number of transactions used in the bulk test. A larger test set produces more statistically robust accuracy measurements. Our quarterly bulk test runs every residential transaction recorded by the Land Registry through our live valuation engine.
Kirchmeyer minimum (50%)
The minimum PE10 regarded as acceptable for mortgage-lending use. Originally defined against surveyor valuations — a softer benchmark than actual sold prices.
Matysiak “acceptable” (70%)
The PE10 level academic research treats as acceptable for routine mortgage lending. Also defined against surveyor valuations rather than actual sold prices.
Rossini “reasonable” (80%)
The international best-practice PE10 target. Like the other thresholds, it was defined against surveyor valuations — a softer test than the sold-price benchmark we use.
Accuracy by segment
Model performance varies by property type, price band, and region. From the Meridian v6.13 H2-2025 bulk test of 295,026 Land Registry transactions.
By Property Type
| Type | n | MdAPE | PE10 | Bias |
|---|---|---|---|---|
| Flat | 48,563 | 7.8% | 60% | +2.7% |
| Terraced | 94,834 | 6.6% | 66% | +2.3% |
| Semi-Detached | 87,931 | 6.2% | 68% | +1.7% |
| Detached | 63,698 | 6.5% | 67% | +1.5% |
By Price Band
| Band | n | MdAPE | PE10 | Bias |
|---|---|---|---|---|
| <£150k | 44,215 | 10.7% | 48% | +14.8% |
| £150-300k | 117,964 | 6.2% | 68% | +1.1% |
| £300-500k | 85,292 | 5.7% | 72% | -0.7% |
| £500k-1M | 41,628 | 6.8% | 66% | -2.0% |
| £1M+ | 5,927 | 9.9% | 50% | -5.3% |
By Region
| Region | n | MdAPE | PE10 | Bias |
|---|---|---|---|---|
| East Midlands | 21,132 | 6.1% | 69.1% | +1.9% |
| East of England | 30,478 | 5.9% | 70.2% | +1.4% |
| London | 37,463 | 6.7% | 66.0% | +2.4% |
| North East | 14,595 | 8.2% | 57.6% | +2.2% |
| North West | 41,147 | 7.5% | 61.1% | +2.7% |
| South East | 45,148 | 6.0% | 70.1% | +1.4% |
| South West | 33,037 | 6.3% | 68.0% | +2.0% |
| Wales | 14,770 | 7.3% | 61.4% | +2.6% |
| West Midlands | 25,190 | 6.5% | 66.0% | +2.3% |
| Yorkshire | 31,967 | 7.3% | 62.0% | +1.9% |
Accuracy by FSD band
Every valuation carries a per-property Forecast Standard Deviation (FSD), mapped to a Fitch RMBS classification band. Accuracy on the current test, stratified by band:
| Band | FSD Range | n | % of test | MdAPE | PE10 | Bias |
|---|---|---|---|---|---|---|
| Band A | FSD ≤ 0.05 | 0 | 0.0% | — | — | — |
| Band B | ≤ 0.10 | 195,609 | 66.3% | 5.9% | 70.4% | -0.6% |
| Band C | ≤ 0.20 | 98,910 | 33.5% | 8.4% | 56.7% | -1.0% |
| Band D | > 0.20 | 507 | 0.2% | 14.7% | 37.9% | +1.8% |
No valuations in the current test cohort carry Band A. The current FSD lookup does not produce FSDs at or below 0.05 for any segment; the band is shown for completeness of the Fitch scale.
Band D valuations show a bias of +1.8% (median signed error) on the current test.
What this means for lenders: For 66% of properties being transacted, our model delivers 70% PE10 with 5.9% median error and -0.6% bias (Band B). The band is reported on every valuation, so you can set your own acceptance policy per band.
Confidence bands
Each property is assigned an individual Forecast Standard Deviation (FSD) based on comparable evidence, feature completeness, and model prediction uncertainty. The FSD maps directly to a Fitch RMBS classification band; the band is the confidence expression on every valuation.
Band A — FSD ≤ 0.05
Forecast Standard Deviation at or below 0.05.
Band B — FSD ≤ 0.10
FSD above 0.05 and at or below 0.10.
Band C — FSD ≤ 0.20
FSD above 0.10 and at or below 0.20.
Band D — FSD > 0.20
FSD above 0.20.
FSD confidence bands
Each valuation also reports a Forecast Standard Deviation (FSD) — a measure of the model’s prediction dispersion for that property. FSD maps directly to the classification bands used by rating agencies such as Fitch for RMBS transactions, giving lenders a standardised confidence signal.
Band A
FSD ≤ 0.05
Narrow prediction interval with dense comparable evidence.
Band B
FSD ≤ 0.10
Strong evidence with reasonable prediction certainty.
Band C
FSD ≤ 0.20
Moderate evidence with wider prediction interval.
Band D
FSD > 0.20
Limited evidence — consider supplementing with desktop review.
Methodology
These metrics are from a quarterly bulk test of 295,026 H2-2025 Land Registry transactions, run through our live valuation engine. They are not training metrics. Every quarter we test our model against every residential transaction recorded by the Land Registry and publish the full results — including known limitations.
Where our model is weaker
Our model performs best on properties in the £150k–£1M range (PE10 65–72%). It is less accurate at the extremes: sub-£150k properties show a +3.8% median bias (overvaluation skew) and £1M+ properties a −5.3% bias (undervaluation). Sub-£150k and £1M+ PE10 sit at ~48–50%, materially weaker than the mid-range. These known limitations are the targets of the v6.13 training-window refresh.
Download reports
Want more detail?
Read our overview for a summary of what we do and how we compare, or dive into the technical summary for methodology, feature categories, and regulatory position.