Swedish asking prices are bait. Gems raises every ask by the district's real asking-to-sold bidding premium (budgivning) before judging value, so a "bargain" that isn't one never reaches you.
The 0–100 Deal Score is deterministic math over price gap, fees, liquidity and risk — never an AI's opinion. Every driver is shown, every valuation is logged and back-tested.
Co-op board (BRF) approval, sublet limits, leased-land (tomträtt) ground rent, non-genuine associations (oäkta föreningar), foreign-buyer friction. Gems knows what kills a rental thesis before you fall for it — and where a freehold house (äganderätt) lets you rent as of right.
— upcoming (kommande) homes are valued against comparable sales before bidding opens — see the fair price before the crowd sets one.
No black box: four steps from raw listing to verdict, with the error rate printed on the tin.
Active listings crawled daily straight from the brokers' own sites (facts only, deep-linked back); 24 months of sold prices per municipality (kommun) and property type for the comps base.
Comparable sales weighted by street, distance, recency and size put a value on every square metre — a model whose error rate is back-tested per market and property type, not assumed. Houses are valued only against houses, and only when the pool localises to their neighbourhood.
Asking becomes expected final price via the district's bidding premium, so teaser pricing can't fake a bargain.
A five-band thermometer from Top deal to Overpriced — and every prediction is logged, so the model's error rate is measured, not claimed.
Honesty rule: with too few comparable sales, Gems widens the window — and if that still isn't enough, it shows a low-confidence range with no verdict rather than a confident guess. Pre-market listings get a modelled range until an asking price exists.