What we do
Asuntokaupat.io prices Helsinki listings with a price model and shows, next to each asking price, the model's estimate, its range and the inputs it was calculated from. The Finder lists the flats asking 5 to 20% less than the estimate.
We report the model's accuracy two ways: its error on listings it never saw in training, and its error against realised sale prices. Both figures are below.
Feedback and corrections: info@asuntokaupat.io or the contact form.
How predictions work
The estimate comes from an XGBoost tree model (v2.5, trained 23 Sep 2026). It is trained on the asking prices of Helsinki listings over 365 days up to 10 Sep 2026, with each flat in the data once even when both listing sites carried it. The data holds 14,523 listings. The model has 27 inputs:
- Size: floor area, room count and floor area per room
- Location: the postcode's price level, the age and number of listings in the area, and distance to the centre, the nearest station and the sea
- Condition: the listing's condition label (good, satisfactory, poor)
- Pipe renovation: status (done, planned, inspection, unknown), year, and how clearly the listing states it
- Building: year built, the flat's floor and the number of floors
- Land: own or leased plot
- Fees: the monthly housing-company fee and the fee per m²
- Amenities: lift, sauna, balcony
The data is split by time: the oldest listings train the model, newer ones validate and test it, and the newest 1,569 are held back until the end. The test then says how the model does on future listings rather than mixing one period into both training and testing.
The homepage table shows each input group's weight in the model.
How accurate it is
Two figures that measure different things:
- On listings the model never saw (1,569 listings, 17 Aug 2026 to 10 Sep 2026): the estimate missed the asking price by 10.5% on average, and half the estimates were within 8.2%.
- On the test set (2,157 listings, 8 Jun 2026 to 4 Aug 2026): average error 11.3%, €38,548 in euros, R² 0.93.
- Against realised sale prices: measured on 28 Apr 2026 with the previous model (v2.4) on 80 Helsinki sales from the Asuntojen hintatiedot service, all closed by the end of 2025. The average error was 17.3% and R² 0.91; 46% of sales landed within ±10% and 70% within ±20%. It cannot be repeated for the current model because the service received no new sales after 2025.
Compare an individual estimate with the realised-sale figure. It includes the gap between asking and sale price and the time from listing to sale, which a test on asking prices does not capture.
Every listing also carries the model's 80% range: the range the model expects 80% of similar flats' asking prices to fall in. When an asking price is below that range, the gap to the estimate is larger than the model's ordinary error.
Before a new model goes live it is compared with the previous one on the same held-back listings. The current model, v2.5, was trained on 23 Sep 2026; the one before it, v2.4, was trained on 28 Apr 2026.
What predictions aren't
Use the estimate as a second opinion: a check on the asking price, a starting point for negotiating, and a way to spot flats priced differently from similar ones. Then visit, ask the property manager, and read the housing company's documents (articles of association, manager's certificate, condition surveys).
Things to watch for:
- The estimate describes an asking price, not a sale price. Realised sales usually closed below the asking price; the listing page shows the figure for its postcode.
- Brand-new buildings and freshly gutted flats sit outside the training data. The model still produces a number, with less to go on.
- Pre-emption clauses (lunastuspykälä) and HITAS regulation are shown when the listing states them, but the model does not price what they cost a buyer.