Methodology
How we get to the number.
No black box. This page lays out the valuation model, the risk checks, the data we use, the confidence framework — and, just as importantly, what we can't tell you and why.
The valuation model.
A hedonic regression — treats the price as the sum of contributions from a property's features, each weight learned from thousands of real Portuguese transactions.
Explains roughly three-quarters of the variance in price per square metre (cross-validated R² ≈ 0.76), with 10.5% median error on held-out data. That beats typical agent estimates and is competitive with bank valuations.
Location
City, parish-level price effects, block-level price memory, distance to centre and coast, and neighbourhood demographics from INE census data.
Size & layout
Interior square metres, bedrooms and bathrooms, terraces, parking, floor level, and construction decade — the basics, measured against their parish-specific premium.
Condition, from photos
Multimodal computer-vision features on listing photos: renovation quality, kitchen and bathroom material tier, facade condition, ceiling height. Distinguishes a recent renovation from a staged one.
Energy & amenities
Energy certificate (A+ through G), elevator, view quality, ocean-view flag, air-conditioning, solar. Each weighted by its measured t-statistic on Portuguese prices.
Competitive environment
Competing listings within 1 km, short-term-rental saturation, median peer price per m² — the market pressure around the property.
Risk and location checks.
A price estimate alone doesn't tell you whether to buy. We layer 30+ checks on top, grouped into six categories — each citing a named public source.
Seismic
Peak ground acceleration at the specific address against Portugal's national seismic zoning. Historical event proximity for high-risk zones.
Source: LNEC · IPMA seismic hazard maps
Flood
100-year flood return periods, proximity to water bodies, and historical flood records.
Source: APA · national flood risk dataset
Fire
Wildland-urban interface exposure, vegetation density, and past fire perimeter proximity — particularly relevant for interior and Algarve properties.
Source: ICNF · historical fire perimeters
Legal
Habitation licence status, Caderneta Predial vs Land Registry alignment, outstanding mortgages and tax arrears, Simplex-era verification gaps.
Source: Conservatória do Registo Predial · AT · municipal records
Short-term rental
Per-municipality AL eligibility and saturation. We track the 308 municipal rule sets and flag containment zones, freezes, and renewal risk.
Source: Diário da República · municipal AL regimes
Connectivity & environment
Distance and travel time to airport, city centre, supermarkets, and schools. Noise exposure bands and air quality where sensor data exists.
Source: INE · IMT · municipal noise maps
Why we use tiers, not percentages.
A confidence percentage like "73%" looks precise but isn't — it's a compound of model uncertainty, data sparsity, and domain assumptions, dressed up as a single number. We prefer tiers that describe what's really happening.
Exact address, full size, bedrooms, bathrooms, construction year, energy rating, and reliable photos — for a property in a well-sampled parish.
Most high-impact fields filled, property in a familiar parish, plausible comparables found.
Core fields present but several high-impact ones missing, or the parish has fewer comparables, or the property is atypical in size.
Minimal inputs, unfamiliar parish, or a property that sits outside the model's training distribution. Use as a rough reference only.
What we're not.
A tool that's honest about its limits is more useful than one that pretends to have none. The short list:
- Our median error is around 10%. On any individual property, you should expect a range — not a single point.
- Extreme properties — unique architecture, very large or very small, land plots, deep rural — sit outside the training distribution and get lower confidence tiers. We say so.
- Some data is not in any public register: condominium debts, pending AGM levies, unrecorded structural issues. We flag what we can't see rather than pretending we can.
- The model is a reference, not a legal valuation. Banks, courts, and tax authorities need a licensed valuer (Avaliador Reconhecido). We don't replace that — we prepare you for it.
- We update features and retrain the model periodically. Old analyses reflect the model version at the time they were run.
The best way to judge the model is to try it.
Run an audit on a property you know well. Check whether the verdict, the risks, and the confidence tier match what you'd expect.