AVM Accuracy Metrics Explained
PE10, MdAPE, bias, FSD, and hit rate — what they measure, how they differ, and why you need all of them to evaluate an AVM properly.
Why one metric is never enough
Every AVM provider quotes accuracy figures, but the metrics they choose to highlight — and the ones they leave out — can paint very different pictures of the same model. To evaluate an AVM properly, you need to understand what each metric measures, what it misses, and why the full picture requires several metrics working together.
All accuracy metrics are derived from backtesting: comparing the model’s estimates against actual transaction prices for a large sample of properties. The error for each property is calculated as a percentage: (estimate − sale price) ÷ sale price × 100.
The various accuracy metrics then summarise this distribution of errors in different ways. Each reveals a different aspect of model performance.
PE metrics: the percentage within a tolerance
PE metrics — PE5, PE10, PE15, PE20 — tell you what proportion of valuations fell within a given percentage of the actual sale price. They are the most intuitive accuracy measures and the most widely quoted in the industry.
How to read PE metrics
PE5 = percentage of valuations within ±5% of the sale price
PE10 = percentage of valuations within ±10% of the sale price
PE15 = percentage of valuations within ±15% of the sale price
PE20 = percentage of valuations within ±20% of the sale price
PE10 is the industry standard headline metric. When someone says an AVM is “70% accurate”, they typically mean PE10 = 70%: seven out of ten valuations fall within 10% of the transaction price.
Meridian PE metrics
147,188 transactions · Q4-2025
Strengths: PE metrics are intuitive, easy to communicate, and directly relevant to lending risk (a lender cares whether the AVM was within 10% of the true value, not about abstract statistical measures).
Limitations: PE metrics are binary — a valuation that misses by 10.1% and one that misses by 30% both count the same (as failures). They also say nothing about the direction of error or the severity of outliers. That is why you need MdAPE and bias alongside them.
MdAPE: the typical error
MdAPE — Median Absolute Percentage Error — is the median of all absolute percentage errors in the backtest. It tells you the “typical” error: half of all valuations had a smaller error, half had a larger one.
Meridian’s current MdAPE is 6.0%, meaning that the typical Meridian valuation is within 6.0% of the actual sale price.
MdAPE is preferred over mean absolute percentage error (MAPE) in property valuation because the median is robust to outliers. A small number of extreme errors — properties that sold at auction, below-market-value transfers that slipped through filtering, or genuinely unusual transactions — can inflate the mean significantly. The median is unaffected.
What it tells you: How close the model’s estimates are to reality for a typical property. A MdAPE under 6% is considered strong for UK residential AVMs.
What it misses: MdAPE ignores direction (over vs under) and says nothing about the tails of the error distribution. A model with a low MdAPE could still have a concerning number of large misses, which is why PE metrics and FSD complement it.
Bias: systematic over- or under-valuation
Bias is the mean percentage error — the average of (estimate − sale price) ÷ sale price across all test properties. It reveals whether the model systematically overvalues or undervalues.
Why it matters: A model with low MdAPE and good PE10 can still be dangerous if it has strong negative bias — systematically valuing properties above their true market value. This creates a hidden risk for lenders, because the collateral is consistently worth less than the AVM suggests.
Conversely, a model with strong positive bias (systematically undervaluing) is conservative but may cause lenders to decline applications that should be approved, or to cap LTV at lower levels than necessary.
The ideal is near-zero bias — the model is neither systematically high nor systematically low. Bias within ±2% is generally considered acceptable for UK residential AVMs.
FSD: property-level confidence
Forecast standard deviation (FSD) is fundamentally different from the other metrics on this page. While PE10, MdAPE, and bias are all aggregate measures — they describe the model as a whole — FSD is a property-level measure that varies from one valuation to the next.
FSD tells you how much uncertainty surrounds a specific valuation, based on how much evidence the model had for that particular property. A standard semi in a well-transacted suburb gets a low FSD; an unusual property in a rural area gets a high one.
FSD is the basis for confidence tier assignment and prediction interval calculation. It is the metric that allows lenders to set differentiated acceptance policies: accept Tier 1, review Tier 2, escalate Tier 3.
For a full explanation of FSD, including interpretation tables and worked examples, see our dedicated FSD guide.
Hit rate: the metric everyone forgets
Hit rate — also called coverage or match rate — is the proportion of properties for which the AVM can produce a valuation. It is the most overlooked metric in AVM evaluation, and one of the most important.
An AVM that declines to value 40% of properties can appear more accurate than one that values 95%, simply because it only attempts the easy cases. The accuracy figures are genuine, but they apply to a narrower, easier population. This is survivorship bias, and it makes like-for-like comparison between AVMs difficult unless hit rate is reported alongside accuracy.
What to look for: Accuracy metrics should always be accompanied by the hit rate. A PE10 of 70% on 95% of properties is more impressive than PE10 of 75% on 60% of properties. The former model is tackling hard cases that the latter model declines.
The practical impact: If an AVM cannot value a property, the lender must fall back to a more expensive alternative (desktop review or physical valuation). A low hit rate increases operational cost and reduces the efficiency gains that AVMs are meant to deliver.
Why segmentation matters
National aggregate metrics are a starting point, not a conclusion. An AVM with a strong overall PE10 might underperform significantly for a particular property type, price band, or region. If that segment happens to be where your portfolio is concentrated, the aggregate is misleading.
The most important segmentation dimensions are:
By property type
Flats, terraced houses, semi-detached, and detached properties behave differently. AVMs typically perform best on terraced and semi-detached homes (high transaction volumes, homogeneous stock) and weakest on detached properties (more variation in size, land, condition).
| Type | MdAPE | PE10 | n |
|---|---|---|---|
| Flat | 7.2% | 63.0% | 23,817 |
| Terraced | 5.9% | 69.3% | 47,321 |
| Semi-Detached | 5.6% | 72.1% | 43,907 |
| Detached | 6.0% | 70.4% | 32,143 |
By price band
Performance often varies with property value. Lower-value properties tend to have tighter percentage errors (smaller absolute amounts at stake), while higher-value properties can be harder to value due to greater heterogeneity and thinner transaction volumes.
| Band | MdAPE | PE10 | n |
|---|---|---|---|
| <£150k | 9.3% | 52.6% | 21,652 |
| £150-300k | 5.6% | 72.0% | 59,783 |
| £300-500k | 5.3% | 75.1% | 43,719 |
| £500k-1M | 6.3% | 68.8% | 20,176 |
| £1M+ | 10.0% | 50.0% | 1,858 |
By region
Regional variation reflects differences in market liquidity, property heterogeneity, and data coverage. Well-transacted regions with homogeneous housing stock typically show stronger AVM performance than areas with sparse transactions or diverse property types.
By confidence tier
This is the most actionable segmentation. It tells you how well the model performs when it says it is confident (Tier 1) versus when it flags uncertainty (Tier 2 or 3). If the tiers are well-calibrated, accuracy should improve significantly from Tier 3 to Tier 1.
For Meridian’s full segmented results, including regional breakdowns, see our accuracy page.
Meridian’s current accuracy
The following metrics are drawn from Meridian’s live backtest against 147,188 transactions across England and Wales.
| Metric | Value | What it means |
|---|---|---|
| PE10 | 69.4% | 69% of valuations within ±10% of sale price |
| PE5 | 43.3% | 43% within ±5% |
| PE20 | 89.4% | 89% within ±20% |
| MdAPE | 6.0% | Typical valuation error |
| Bias | +1.6% | Systematic over-valuation tendency |
| Sample size | 147,188 | Transactions in back-test |
These figures are updated regularly and are available with full segmentation on our accuracy page. Lenders can also run their own backtest using their portfolio data via our free backtest service.
Verify these numbers yourself
Upload your own transaction data and see Meridian’s metrics on your portfolio. Free, unlimited, no credit cost.