A home in the register
Registered homes — we know the year built and official assessment.
about 9 in 10 within 20%; the typical miss is 5.5%.
Measured on 6,171 out-of-sample sales.
Listing fields only — simulated
Live listings — the model doesn't use year built or assessment as inputs (the official assessment may still be shown for reference), so it relies on size, rooms, type, the exact location and the local price level around it.
This test masks the fields a listing does not carry — build year, age and the official assessment — on completed register sales. It is not yet an evaluation of real adverts matched to their later sale price.
about 8 in 10 within 20%; the typical miss is 6.5%.
Measured on 6,369 out-of-sample sales.
Older sales are adjusted to a comparable price level using the official house-price index before training. This helps the model compare sales from different periods without confusing location, size and general market appreciation.
Where the model is strong — and where it isn’t
The same backtest, broken down. Accuracy varies by property type, data coverage and price band; each breakdown is measured on its own, and the ones available are shown below.
"Sale price inside range" is measured on the last test period (Jul 2026 – Oct 2026), with ranges fitted only on the earlier periods.
A home in the register
Floor area as recorded in the property register, not the advertised area.
Listing fields only — simulated
Floor area as recorded in the property register, not the advertised area.
What the estimate can’t see
Every estimate is a range, not a guarantee. The model values a home from its size, rooms, location, type and (for registered homes) year built — it can’t see condition, renovation, floor, view or parking, which is why we always show a range and a confidence level rather than a single false-precision number.
Methodology & metric definitions
Measured out-of-sample on real sold prices (rolling backtest): for each metric the model was trained only on sales before the test window. PE10/PE20 = share of estimates within 10%/20% of the eventual sale price; MdAPE = median error; COD = dispersion (≤15 is professional-grade); PRD ≈ 1.0 is fair across price levels (>1.03 = mildly under-pricing expensive homes); 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%). Per-segment accuracy comes from the same held-out predictions; a segment with fewer than 150 backtested sales is flagged as limited data, and a postcode counts as well-covered at 200+ register sales. Each segment's range coverage is measured on the same held-out fold and quoted only when the segment has at least 100 held-out sales. The price band groups by the model's OWN estimate, not the actual price. Estimates are a RANGE, never a guarantee — the model can't see condition, renovation, floor or view.
Model trained Oct 2026 on 161,673 sales. Source: HMS Kaupskrá (sold prices) — leakage-free rolling backtest.
This model: commit 4aa7965150 · artifact sha256:cc7377429d62e8f6… · data through 2026-10-02 · 161,673 sales used · 0 implausible-price rows excluded