Methodology & data sources

The honesty contract

Every analytics response — anything computed, scored or aggregated — carries a data_quality flag: real (computed from a live official source), partial (some inputs were unavailable or cached; the note says which), or synthetic(placeholder, never presented as data). Raw-lookup endpoints (address search, a property's transaction list) return register rows verbatim and carry no flag — there is nothing modelled in them to caveat. When a source is down or a sample is too small, the figure is omitted, not estimated: missing beats misleading. Provenance notes ship in the payload itself, so the caveats can't drift away from the numbers.

Data sources

Listing data is sanitized before display: implausible bedroom counts (more bedrooms than total rooms), missing bathrooms, rental-priced “sales” and duplicate entries are dropped or nulled rather than shown. The Icelandic room convention is kept: herbergi = bedrooms + living rooms; kitchens and bathrooms are not in that count.

The valuation model

A single gradient-boosted model predicts the (log) sale price of a home, trained on the full register. Every home is placed at its exact coordinates (from Staðfangaskrá) and the model also sees the local price level — the median price/m² of sales within roughly a kilometre over the prior 24 months — so location matters at street level, not just postcode level. For homes in the register it additionally sees size, rooms, dwelling type, build year and official assessment; for live listings the model does not take build year or assessment as inputs (the official assessment may still be shown for reference), which is why the two are reported as separate accuracy regimes above.

Accuracy is measured the only honest way: a leakage-free rolling backtest — four folds, each trained strictly on sales before its test window, scored on the sales after it (a sale can never inform its own features, either). We publish the median error (MdAPE), the share of estimates within 10% and 20% of the eventual sale price, the coefficient of dispersion (COD — ≤15 is the professional-assessment benchmark), the price-related differential (PRD ≈ 1.0 means it is fair across price levels), and the range coverage — how often the displayed low–high range contained the actual sale price, measured on a held-out fold the ranges were never fit on (target ~80%). The full table is on the accuracy page.

Every estimate is shown as a range, never a point. The model cannot see condition, renovation, floor, view, parking or garden — a renovated penthouse and a fixer-upper of the same size in the same building get the same estimate. That is a disclosed limitation, not a bug to paper over.

The listing deal score

A listing's asking price is compared against the modelled value (or, when the model can't price a segment, against the median sold price/m² of comparable register sales: same postcode, same dwelling type, ±30% size, trailing 24 months — the comparables are shown on every listing page). “Below modelled value” is a candidate worth investigating, never a verdict: a big discount usually means the market knows something the model can't see.

Affordability

Two complementary reads, each used where it is defensible:

• Price-to-income multiple (Demographia convention): median sale price ÷ gross income of a modelled two-median-earner household. This anchors the national affordability score (≤3× affordable … 5×+ severely unaffordable).
• Greiðslumat payment burden(Iceland's actual standard): the mortgage payment as a share of take-homehousehold income, against the Central Bank's 35% cap (40% for first-time buyers). Shown for both mortgage products — non-indexed (óverðtryggt, priced off the live policy rate + 1.0pp retail spread) and CPI-indexed (verðtryggt, priced off the live indexed par yield + 1.3pp). The indexed figure is the initial payment and rises with inflation — a cash-flow timing benefit, not a lower real cost. The affordability score never uses the indexed teaser payment.

Affordability verdictsexist only at the national level. Postcodes and small municipalities are income-sorted, so stamping a neighbourhood “affordable” or “in crisis” from its area median income is an ecological fallacy — for small areas we show transparent figures (payment, burden vs the caps) and let you judge.

Matching listings to the register

When a live listing can be matched to its register entry, its own sale history and official assessment are shown. The match rule is conservative: a per-unit claim is made only when exactly oneregister unit at that address fits the listing's size; ambiguous cases degrade to a clearly-labelled building-level view with no per-unit valuation, and no match shows nothing. Administrative deeds (symbolic 1,000-ISK transfers) are excluded from sale histories so they never masquerade as market prices.

Sample-size rules

Aggregates publish only above minimum sample sizes — for example a municipality needs 50+ residential sales in the trailing year before it gets its own affordability row, and quarterly postcode trends require a minimum number of sales per quarter. Below the threshold the figure is withheld and the gap is the message.

Known limitations

• Prices are nominal unless a chart says otherwise; long histories are not inflation-adjusted in tables.
• The register records registered deeds — recent weeks are always incomplete.
• Income data (TEK01002) lags by roughly a year; affordability uses the latest published year, stated on the page.
• Listing coverage is Vísir's, not the whole market; bed/bath fields from the portal are unreliable and are sanitized.
• No rental-market data yet (the public rent registry has known revision issues; we'd rather wait than publish numbers that get revised under you).

Corrections

Spot something wrong — a number, a label, a caveat that undersells the uncertainty? Tell us. Being correctable in public is part of the method.