How PAID values properties

Every property valuation on PAID is one of four kinds. Each comes with a range, a confidence indicator, and the source it was derived from — because a single number on its own hides where it came from and how much you can trust it.

Where does this number come from?

We use four sources, in order of precision:

1. Zoopla AVM (UPRN match)

When we can match the exact property in Zoopla's automated valuation database via UPRN (Unique Property Reference Number), we use Zoopla's own AVM ( sale_estimate_current). This is the most precise source — Zoopla's AVM is informed by the specific property's history, condition signals, and recent local sales.

Pill label: Zoopla. Confidence: usually very high. Range: Zoopla's own published lower / upper bounds.

2. Zoopla AVM (area match)

When no UPRN match exists, we fall back to the nearest Zoopla AVM within 50 m via PostGIS spatial query. The number reflects the immediate area but isn't tied to the exact property, so we flag this distinction in the pill.

Pill label: Zoopla (area). Confidence: high to medium. Range: Zoopla bounds, possibly widened by the confidence-tier floor.

3. PAID Estimate (comparables analysis)

When neither AVM tier produces a value, we run our own comparable-sales analysis. The pipeline:

1. Fetch comps from Land Registry Price Paid Data, Zoopla sold prices, and Rightmove sold listings within a radius (typically 0.25–1 mile) and a recency window (24 months).
2. Score similarity on type, bedrooms, bathrooms, street, postcode, distance, recency, square footage, EPC band, and construction-age band. Comps below a quality threshold of 60/100 are dropped.
3. Time-adjust historical comp prices to today via the UK House Price Index (HM Land Registry, postcode-area regional level). A flat sold for £200k three years ago in an area where the index moved from 100 to 110 is treated as a £220k comp.
4. Weight + average. Each surviving comp is weighted by its similarity score; we take the weighted mean as the central value.
5. Derive a range from the inter-quartile range (IQR) of comp prices, widened to a confidence-tier floor (very_high ≥ ±3%, high ≥ ±5%, medium ≥ ±8%, low ≥ ±12%). The floor protects against falsely-narrow ranges when 5 comps happen to share a price.

Pill label: PAID Estimate. Confidence: depends on comp count + dispersion. Range: derived as above.

We deliberately do not apply a blanket conservatism discount to the central value (older versions of our pipeline applied a flat 3%). Conservatism lives in the lower bound of the range now — that's a more honest signal.

4. Manual override

If a user sets a fair-value override on a deal (e.g. a buyer adjusts our number based on their own viewing notes), the manual value takes precedence. We still record what our underlying estimate was at the time so the user can see the gap.

Pill label: Manual. Confidence: not applicable.

How reliable is the figure?

Every valuation surfaces a confidence pill colour:

• Green (very_high or high): trust the number for decisions; range is tight; comp set is rich.
• Amber (medium): plausible but worth a sanity check against the comps tab; range is wider.
• Grey (low): treat as a wide guide; run your own comps before you act.

The range band itself encodes precision. A green pill with a tight band (±3%) is a much stronger signal than a green pill with a wider band (±5%) — both are "trustworthy" but one is more precise.

What is it based on?

For comp-based estimates, hover the pill to see:

• How many comparable sales fed the calculation
• The radius they were drawn from
• How recent they were (months)
• Whether the range came from comp dispersion (IQR) or the confidence-tier floor

For Zoopla AVM tiers, the tooltip names the match type (UPRN-exact vs area-50m).

Limitations + caveats

• Estimates assume condition is "average for area". A property in unusually good or bad condition may sit outside the range. We're working on a condition-aware adjustment for a future release.
• Time-adjustment uses the regional UK HPI (12 regions). A specific micro-market may move differently — e.g. an outer-London suburb with a new tube extension might outperform the South East regional index.
• Our backtest currently produces an upper-bound MAE of ~33% on v1 snapshots when joining LR sales to properties on postcode alone (without street/PAON match). Tighter measurement is in progress; treat the backtest as a relative-comparison signal, not an absolute accuracy claim.
• We are not RICS Red Book compliant. PAID estimates are decision-support tools for property investors, not regulated valuations for mortgage lending. If you need an RICS valuation, commission one.

Versioning

Every BMV calculation is stamped with a methodology version (currently v1 for legacy rows pre-April 2026, v2 for new rows). The version is visible on the methodology tooltip for any specific property. When we change the comp scoring, comp sources, or range derivation, we bump the version so older recommendations stay traceable.

Last updated: April 2026. Methodology version: v2.