duvet · Model summary
Custom-trained property intelligence.
Custom-trained EPC and valuation models, measured against homes excluded from training. The consumer edition works with the details you know and shows calibrated estimate ranges.
01 · Construct
Learn from real buildings
We train gradient-boosted tree models on recorded EPC characteristics and residential sale prices in England and Wales. The models learn relationships between location, size, construction, heating, energy performance and value.
02 · Estimate
Work with what you know
The models use size, construction, glazing and heating details. Valuation adds location and aggregated sale-price context. Unknown wall and heating details remain unknown, with their own tested ranges. The energy model estimates a certificate score and maps it to an indicative EPC band.
03 · Evaluate
Test beyond training
Each property appears in only one dataset split. Training, model tuning, range calibration and final testing are separate. Valuation is tested on later sales in 2025 and January 2026; geographic context excludes every held-out property.
Consumer edition · 6 October 2026
Current valuation benchmarks
Recorded sale prices for 5,059 fresh homes, sold after the training period. An earlier diagnostic test was quarantined after a location-feature defect; these homes were reserved before evaluating the corrected model. Local sale context uses training properties only; training features exclude the property’s entire fold.
Scroll sideways to see all metrics. Input labels stay visible.
| Inputs | Test homes | Median error (MdAPE) | Mean error (MAPE) | Within ±10% | Within ±20% | MAE |
|---|---|---|---|---|---|---|
| Complete details | 5,059 | 11.54% | 17.09% | 44.14% | 72.19% | £55,872 |
| Wall and heating unknown | 5,059 | 11.70% | 17.30% | 43.92% | 71.91% | £56,339 |
The 90% range covered 90.93% of complete-input test prices and 90.87% with unknown walls and heating. Market context uses actual type-specific HPI prices through July 2026, with a three-month lag. HPI history is the revised July 2026 snapshot, rather than an unrevised historical backtest. This tests recorded sales, rather than current asking prices or surveyor valuations.
Consumer edition · 6 October 2026
Current EPC benchmarks
A separately trained eight-input model, tested on 19,988 distinct homes with public-equivalent details. The target is an observed certificate score from sold-home records.
Scroll sideways to see all metrics. Input labels stay visible.
| Inputs | Test homes | MAE (score points) | Within ±10 points | Exact EPC band | 90% range coverage |
|---|---|---|---|---|---|
| Complete details | 19,988 | 5.75 | 83.84% | 60.78% | 90.04% |
| Wall and heating unknown | 19,988 | 6.72 | 80.07% | 58.05% | 90.06% |
Bedrooms, fuel and hot water are excluded because equivalent training fields were unavailable. Wall and heating unknowns were deliberately tested and calibrated separately. Ranges describe observed test coverage; individual homes can fall outside them. A predicted band requires an accredited assessment to become an official EPC.
Historical research models and reported benchmarks
Internal valuation evaluation · 17 April 2026
v28 benchmark results
Historical grouped-UPRN validation report, currently under reproduction review. Results are from the report’s 30,000-tree model configuration, evaluated against recorded sale prices.
| Property type | Validation records | Median error (MdAPE) | Within ±10% | Within ±20% | MAE |
|---|---|---|---|---|---|
| Flat | 95,489 | 9.2% | 53.3% | 80.2% | £38,750 |
| Semi-detached | 167,983 | 8.39% | 56.9% | 83.8% | £29,244 |
| Terraced | 167,821 | 9.27% | 53.1% | 79.8% | £30,085 |
| Detached | 147,942 | 9.64% | 51.4% | 79.1% | £56,920 |
Source: duvet v28 benchmark report, generated 17 April 2026. The source report calls the median metric “median MAPE”; we label it MdAPE here to distinguish it from mean percentage error. The audit found category-encoding differences between training, the benchmark and serving, along with feature-unit and routing defects. The recovered ONNX specialists contain 10,000 trees, while this report uses 30,000. These figures are reported historical validation results, not independently verified current-demo accuracy. Validation also selected model stopping points.
Historical EPC research · 20 March 2026
Energy prediction performance
4.49
SAP points mean absolute error
633,278
Cleaned records: training and validation combined
Historical enhanced-input experiment with 24 features: 531,441 training records and 101,837 geographic validation records. Validation also selected the stopping point. This run did not load PPD enrichment despite its name. Serving input differences remain under review, so these results do not establish accuracy for the simplified demo or current service.
Source: duvet EPC training handover, G4_PPD experiment completed 20 March 2026. A predicted band is an indicative estimate, not a statutory EPC.
Understanding the numbers
Accuracy metrics explained
- MAE
- Mean absolute error: the average size of the prediction error, expressed in pounds for valuation or SAP points for energy.
- MdAPE
- Median absolute percentage error: half of evaluated predictions have a smaller percentage error, and half have a larger error.
- MAPE
- Mean absolute percentage error: the average percentage error. It can be more affected by large errors than the median.
- Within ±10% / ±20%
- The share of predictions within 10% or 20% of the reference value. These are measured hit rates, not an accuracy percentage for an individual home.
Market context · Sources checked 2 October 2026
Published industry benchmarks
Providers use different markets, reference values, coverage and evaluation periods. These public figures provide context; they are not a common-dataset test or evidence that duvet outperforms another provider.
Models can miss condition, recent alterations and unusual properties. An indicative valuation does not replace a surveyor’s formal valuation, and an energy prediction does not replace an accredited assessor’s certificate.
← Try the modelsContains HM Land Registry data © Crown copyright and database right 2026. Contains OS data © Crown copyright and database right 2026. Public sector information is used under the Open Government Licence v3.0. Source permissions for a paid production API require a separate review.
Property intelligence at scale.
Planning a portfolio analysis or product integration? Talk to us about EPC predictions, automated valuations and API access.