How it works

How EPC Benchmark builds a comparison

This page sets out exactly what data this demonstration uses, how the comparison group is built, what the figures mean and where the method stops being reliable. It also covers privacy and accessibility.

Methodology

1Keep usable domestic recordsStart from the analysed Manchester domestic certificate export and keep records with the fields needed for a fair comparison.
2Select the latest assessmentFor each UPRN, keep only the most recent usable assessment in the analysed export. That record is what we call the “latest” one.
3Match on recorded attributesMatch on local area, property type, built form, age band and floor-area band, using only fields recorded on the certificate.
4Widen only as neededUse the narrowest reliable cohort first, then widen one attribute at a time until the group reaches the target of at least 30 comparable latest certificates.
5Compare the scoreCompare the home's score with the cohort median and work out an approximate percentile position within that group.
6Join recorded improvementsJoin recommendation rows by certificate identifier, preserving the recorded sequence, indicative cost ranges and improvement IDs.

Data notes

Certificate rows analysed: 268,839; Domestic certificate rows in the analysed Manchester export.
Unique UPRNs: 192,142; Distinct property identifiers present across those rows.
Usable latest property records: 153,049; One record per property: the most recent usable assessment for that UPRN.
Cohort coverage: 96.9%; Share of usable latest records that could reach a comparison group of at least 30. A dataset coverage figure, not a performance claim.

Improvement measures come from a separate recommendations export, of which only 25,000 rows are used in this demonstration. Recommendation rows are joined to a certificate by its identifier and shown in the recorded sequence, with the recorded cost range and improvement IDs preserved. Because the sample is row-limited, the measures shown for any certificate are not guaranteed to be its complete recommendation package.

Cost figures are indicative historic EPC assessment ranges. They are not quotations, not current market prices and not adjusted for inflation.

Limitations

EPCs are modelled assessments and can become outdated as homes change.
Comparability depends entirely on the fields recorded by the assessor.
The percentile position is approximate, not an exact ranking.
The recommendations data used here is only a 25,000-row sample, so a certificate's full package may include more items.
Cost ranges are indicative historic EPC assessment ranges, not quotes or current market prices.
This is an independent demonstration, not an official register and not a professional survey.

Privacy

This demonstration does not collect or store a submitted address, because it only uses one anonymised example property. There are no accounts, no sign-in, no database, no payment processing, no advertising and no analytics tracking. Nothing you do on this site is sent anywhere for storage.

No street address, UPRN, certificate number, API token or private certificate record is included in this release.

Every control is reachable and operable by keyboard, with visible focus outlines.
Semantic headings, lists and landmarks, plus a skip link to the main content.
Band and status information is always labelled in text, so nothing depends on colour alone.
The comparison chart has a full text alternative describing every plotted value.
Touch targets are at least 44 pixels, and motion is reduced when the system asks.

Next release

A later version can connect securely to official EPC data, with credentials held and used server-side only. An API token must never be placed in browser code, so any live lookup would run through a protected server-side request rather than from the visitor’s device. Until then, this release stays a clearly labelled demonstration with one anonymised example.