How Alpine Market builds a property record
What we read, how duplicate listings are merged, what a canonical property record is, how the Alpine Estimate is computed and when it is withheld, where recorded sales come from, and what we never publish.
How listings are aggregated and deduplicated
Each portal we can reach is read on a rotation. Every appearance of a listing is stored as its own row with the payload we read, so a record can always be traced back to what the source actually said on the day we read it. Those rows are never deleted.
Same reference on the same portal
An appearance that carries a reference we have already seen on that portal updates the existing record rather than creating a new one.
Same address
Addresses are normalized (street forms, abbreviations, accents, postal codes) and matched against the federal building register, so two portals writing the same address differently still land on the same building.
Scored match on everything else
What falls through is scored on distance, property type, rooms, floor area and price. A pair has to be within 75 metres, the same type, the same room count, within 5 percent on area and within 3 percent on price before it can merge automatically. A signal missing on either side scores zero; it is never quietly dropped from the calculation.
Photographs
Listing images are hashed, both exactly and perceptually, so the same photograph recognises itself across portals even after it has been resized or recompressed.
Pairs that score high but do not clear the automatic rule go to a review queue rather than merging on a guess. Every merge is recorded with the evidence that produced it and can be undone.
The canonical property record
A property is the physical asset: a building, or a unit inside one. A listing is a marketing event about that asset, and a source row is one portal's version of that event. Keeping the three apart is what lets a record survive a relisting. When a flat comes off the market and reappears eight months later at a different price through a different agency, it is the same property record, now with two listings and a price history, rather than a second property with no past.
Fields written by our collectors are refreshed on every read. Fields confirmed by a person, or produced by our own enrichment, are preserved rather than overwritten by a thinner source.
How the Alpine Estimate is computed
The Alpine Estimate is a comparables model. For a given property we assemble comparable properties in the same canton, in the same valuation cohort as the subject, last seen on their portal within the last 120 days, and compare on price per square metre. The cohort is broader than the exact property type: apartments, studios, duplexes, penthouses, attic flats, lofts and furnished flats are one cohort, and houses, villas, chalets and terraced houses are the other, so a studio can be valued against an apartment but never against a chalet. A property type in neither cohort is not valued at all. Implausible values and statistical outliers are trimmed before anything is computed. The subject property is excluded from its own comparable set.
The published range is the tenth to ninetieth percentile of that comparable band, not a decorative margin around a point estimate. A narrower band was tested and rejected because it did not contain the outcome often enough to be worth showing.
The model fails closed. Fewer than five eligible comparables, a missing floor area, a property type the model has not been calibrated for, or a calibration run that did not meet its accuracy and coverage gates all produce a withheld estimate with the reason stated, not a number with a shrug attached. Accuracy is measured by leaving each observation out of its own comparable set and recalibrating, and the result has to clear a fixed error and coverage threshold before an estimate is shown at all.
The most important limitation, stated plainly: today the evidence is disclosed asking prices from live listings, not recorded sale prices. An Alpine Estimate is therefore an estimate of what a property would be asked for in the current market, and it is not a valuation for lending, tax or legal purposes. Swiss recorded-price coverage is thin enough that folding it in would bias the model rather than improve it; when that changes, this page changes with it.
How recorded sales are sourced and linked
Recorded sales are transfers published by an official body, distinct from asking prices. We read the Geneva Feuille d'avis officielle for Switzerland, the DVF open dataset for France, and Price Paid Data from HM Land Registry for the United Kingdom. Counts and dates for each are on the data sources page.
Linking a recorded transfer to a specific building is a separate problem from reading it. A register entry describes a parcel and a commune in the register's own vocabulary, which does not always resolve to one address. Where we cannot resolve it with confidence, the record stays in the registry and is counted there, but it does not appear on a property page. Most Geneva records are in that state today, and the data sources page publishes the measured share rather than implying full linkage.
Not every recorded transfer is a market price. A sale at auction, a forced sale, an exercised pre-emption right or a transfer settled in kind is a real transaction, so it is listed, but its price is set by a special mechanism rather than by an open market, so it is excluded from medians, quartiles and comparable sets.
What we withhold, and why
Names of private individuals
No name of a natural person is published anywhere on this site: not a buyer, not a seller, not a notary, not an agent. Recorded transfers are published as price, date and type. This is a fixed rule, not a default setting, and it is enforced in code rather than left to the care of whoever writes the next page.
Transfers whose nature is unclear
An inheritance, a gift and a purchase can look similar in a register notice, and the price attached to a transfer that is not a purchase says nothing about market value. When the source does not make the nature of a transfer clear, we withhold the price and do not treat the record as a sale. The record is kept, so it can be revisited if better evidence arrives, but it is never published as a transaction on a guess.
Register addresses we hold nothing for
Address search reaches every building address in the federal register, far more than we hold listings for. A page for an address we know nothing else about is a thin page, so those pages are excluded from our sitemaps and marked not to be indexed. They exist for a reader who searched for that address, not for a search engine.
Found something wrong?
A wrong merge, a stale listing, a price that does not match the source, or a record you want removed: write to info@alpinemarket.ch and tell us which page. Corrections are worth more to us than any single listing.