Skip to content
Building performance

Modelled versus measured: the housing energy performance gap

What SAP and RdSAP actually model, what RdSAP 10 changed, and what measured studies show about the gap between EPC ratings and real energy use.

An EPC is a model output. That is not a criticism — it is the design. But if you are allocating retrofit spend from EPC bands or a stock model built on them, it is worth knowing precisely what the model assumes, where the measured evidence departs from it, and by how much.

What SAP and RdSAP calculate

Reduced Data SAP (RdSAP) is the approved methodology for assessing existing dwellings and generating their EPCs. The RdSAP 10 specification, written by BRE on behalf of government, states that energy performance is expressed as an "asset rating" — "a numerical indicator of the cost of energy derived from the energy needed to meet different needs associated with standardised use of building".

Standardised use is the point. Occupancy, heating hours and set points are fixed by the method, not observed. Two structurally identical homes receive the same rating whether one is heated to 22°C by a family of five or one room is heated by a single occupant. The rating compares fabric and services. It was never a forecast of a household's bill.

RdSAP also differs from full SAP in how it gets its inputs. It is "a system of data collection together with defaults and inference procedures". Dwellings fall into age bands A (before 1900) to M (2023 onwards), and the age band supplies what the assessor cannot see. Solid brick walls "as built" are assigned 2.5 W/m²K up to 200 mm thick, 1.7 for 200–280 mm, 1.4 for 280–420 mm and 1.1 above that. Where wall thickness itself cannot be measured, it too comes from a table indexed on age band and wall type.

The evidence rule runs one way. Insulation "beyond what would normally be assumed for the age band" can only be specified "if adequate evidence exists". An undocumented retrofit is, to the model, absent.

What RdSAP 10 changed

RdSAP 10 came into use for existing-dwelling EPCs in England and Wales from 15 June 2025, replacing RdSAP 9.94 and running on the SAP 10.2 calculation core.

The direction of travel is towards observation. Every window is now measured rather than inferred from age and built form. Lamps are counted rather than fittings. Recorded insulation thicknesses and thermal conductivities can be entered where evidence supports them. Air pressure test results are accepted, ventilation options are expanded, and Room-in-Roof handling is more granular (Elmhurst Energy summary).

That is a real improvement, and DESNZ expects measured glazing areas to correct a known solar-gain error. It does not change the underlying position: where evidence is absent, the age band still decides.

What the measured evidence shows

Few et al. (2023, Energy and Buildings 288) compared EPC-modelled and smart-meter-measured primary energy use in 1,374 gas-heated British homes from the SERL Observatory. EPCs over-predicted by 8% for band C properties, rising to 48% for bands F and G. The gap persisted in homes that matched the model's assumed occupancy, thermostat set point and whole-home heating.

The DESNZ EPC Accuracy Research (published 26 May 2026, carried out by UCL Energy Institute and Alan Pither Ltd, 1,136 monitored homes) found gas-heated homes' EPCs overestimate delivered energy by 16.0% on average. Correcting for the latest RdSAP version, actual weather, post-EPC upgrades and real occupancy narrows it to 10.9% — and more than half of that improvement comes from measures installed after the EPC was issued. Electrically heated homes use about 31% less than predicted, widening to nearly 47% in December. Assessor error alone shifts predicted space and hot water heating by around 6%.

Critically, the gap is not one-directional. The same report finds heat loss is over-predicted most in older, poorly insulated homes — and reverses in new homes, which lose more heat than modelled.

Both ends are documented. Li et al. (2015, Building Research & Information 43:2) measured in-situ heat flux across 40 brick and 18 stone solid walls and found a mean U-value of 1.3 ± 0.4 W/m²K against the 2.1 then assumed. Applied across the stock, that single change cut modelled mean annual space heating demand by 16% and moved roughly one third of solid-wall dwellings up an EPC band. Their conclusion matters more than the number: "the distribution of U-values is so large that the on-going use of a single mean cannot be justified when assessing individual properties."

At the other end, Johnston, Miles-Shenton and Farmer (2015) coheated 25 new-build dwellings built to Part L1A 2006 or better and measured whole-building U-values around 1.6 times the design prediction.

Why the gap exists

Assumed rather than surveyed construction. Default U-values applied as single values where the real population is a wide distribution. Standardised occupancy, a two-zone heating model the DESNZ data does not support, and a minimum 0.5 air changes per hour that exceeds many measured homes. Workmanship, thermal bypass and as-built variance in new construction. Assessor interpretation. And staleness: EPCs are valid for ten years and do not reflect anything done afterwards.

Why this matters operationally

Because the modelled number gates money. The Warm Homes: Local Grant policy guidance is explicit: funding "must only support households living in England in homes with an EPC rating within bands D to G — this is an Energy Efficiency Rating (EER) SAP score of 68 or below". Bands A to C are ineligible in all circumstances.

Coverage compounds it. ONS analysis (28 October 2025) found 71% of English dwellings and 67% of Welsh dwellings have had at least one EPC since 2007, but only 61% of English pre-1930 stock. The homes with the thinnest data are the homes where age-band inference is least reliable.

The operational consequence is ordinary. Surveys get sent to addresses chosen by age band and built form. Some of those homes are already better than the model says. Others — newer, banded C, ineligible — are worse. Change one assumption and a third of a stock category moves a band.

What measurement adds, and what it does not

Thermal observation shows the building as it is rather than as its age band implies. A pre-1900 terrace with an undocumented internal wall insulation retrofit looks identical on paper to its neighbours. In the infrared, it does not.

Thermal imaging has real limits, and they are not small. It measures apparent surface temperature, not U-value. BS EN ISO 6781-1:2023 frames the technique as detection of heat, air and moisture irregularities — qualitative by design. Readings are sensitive to indoor–outdoor temperature difference, wind, solar loading earlier in the day, thermal mass, surface emissivity, reflected temperature, viewing angle and recent rain. Reliable work wants night or pre-dawn capture, a sustained temperature difference across the envelope, and low wind. It cannot see through render, cladding, foliage or a parked van, it cannot see the elevation facing away from the street, and it cannot see inside a cavity.

It narrows where to send an expensive survey. It does not replace one. The same questions about provenance apply to a thermal reading as to any other.

Where Make Sense sits

Make Sense is building towards street-level thermal measurement, using sensors mounted on vehicles already making their rounds. The first vehicle system is in build. Nothing is deployed, and there are no accuracy or performance figures to quote, because nothing has been validated. When there are, they will be published alongside their limits and their provenance.

If you are targeting retrofit from modelled stock data and want to talk about where measurement would and would not help, get in touch.

Sources

Last reviewed 12 August 2026

Related
Talk to us about this