Methodology & measured accuracy
Most property estimates are published without any statement of how wrong they can be. Below is how DwellIQ produces a value, and how far its estimates sat from the price properties actually sold for, measured out of sample.
Measured accuracy
The model was fitted on 2023 recorded sales and then tested against 2024 sales it had never seen, across 18 communes spanning dense urban, mid-size and rural markets. Every transaction in the test set is a real recorded sale, so no estimate is graded against another estimate.
Median error
19%
Half of estimates fall within 19% of the recorded sale price. Measured on 35,828 transactions.
Within 20%
52.1%
52.1% of estimates landed within 20% of the sale price; 28.1% within 10%.
Mean error
30.5%
Higher than the median because a minority of atypical properties are badly mispriced by any surface-based model.
Systematic bias
+3.6%
The typical estimate sits slightly above the eventual sale price, partly because the most recent published sales predate the listing.
By market type
| Market | Median error | Within 20% | Sales tested |
|---|---|---|---|
| dense urban | 16.7% | 57.5% | 14,254 |
| mid-size city | 20.6% | 48.8% | 19,541 |
| small town / rural | 21.4% | 47.3% | 2,033 |
| Appartement | 18.3% | 53.6% | 29,385 |
| Maison | 22.4% | 45.4% | 6,443 |
Accuracy is best in dense urban markets, where there are many recent comparable sales, and worst for houses in thin rural markets, where individual plots, land area and condition dominate the price.
How the estimate is produced
- Recorded sales, not asking prices. Comparable values come from DVF, the French tax authority’s register of actual property transactions. Multi-lot sales are excluded, because their single declared price cannot be attributed to one dwelling, as are transactions implying an implausible price per m².
- A size curve, not a flat average. Price per m² falls as properties get larger, so applying one commune average to every property over-values large homes and under-values small ones. We fit price against living area on the commune’s own sales. Replacing the flat average measurably reduced median error from 19.9% to 19% and cut systematic over-valuation from 5.7% to 3.6%.
- A range, not a single number. The published range is the interquartile spread of the model’s own residuals on that commune, so it widens where the local market is genuinely harder to read.
- Rent from official indicators. Rental income uses the government’s per-commune rent indicators, selected by property type and size band, with the published confidence interval carried through into the yield range.
- Deterministic scoring. Scores and the resulting recommendation come from a fixed model: the same inputs always produce the same score. Language models are used only to write the explanation, never to set a number.
Known limitations
These figures are what a surface-and-location model can achieve on open data alone. The estimate does not know a property’s floor, aspect, state of repair, outdoor space, parking, or the works it needs — factors that routinely move real prices by more than 20%. Recorded sales are also published with a lag, so the most recent comparable data predates a live listing by some months.
This is why DwellIQ is a screening tool: it is well suited to ranking and filtering properties quickly, and not a substitute for a valuation by a professional who has stood inside the building. Estimates supplied with the property’s own characteristics — floor, condition, works — are materially more accurate than estimates derived from a listing alone.
Sources
- DVF — Demandes de valeurs foncières, © DGFiP, published on data.gouv.fr under Licence Ouverte 2.0.
- “Carte des loyers” rent indicators — Ministère de la Transition écologique / ANIL, published on data.gouv.fr under Licence Ouverte 2.0.
Fitted on 2023 sales, tested on 2024. Accuracy is recomputed when the model changes; the test harness is part of the codebase.