How Estiq calculates property values across Estonia
Estiq's Automated Valuation Model (AVM) is a statistical tool that produces an instant estimate for any flat in Estonia. In Tallinn it compares the subject with similar flats nearby, adjusts each comparable's price to the subject's characteristics and brings asking prices to transaction level using Land Board (Maa- ja ruumiamet) sales statistics. The result is a price with an 80% range; the model's accuracy is measured with a blind test (section 5).
1. What the AVM is — and what it is not
Estiq's Automated Valuation Model (AVM) is a statistical tool that produces an instant property price estimate for any Estonian address. It compares the subject with similar nearby flats, adjusts for size, construction era, condition, energy class and floor, brings asking prices to transaction level using Land Board statistics, and outputs a price with an 80% range.
This is not an EVS 875 certified appraisal
Estiq's AVM is a statistical estimate. It does not include a physical inspection of the property. Estonian banks require a certified appraisal under the EVS 875 standard — produced by a licensed hindamisekspert after an on-site visit — as mandatory collateral documentation for mortgage lending.
Use Estiq's AVM for: pricing strategy, market research, buyer offer benchmarking, broker presentations, and portfolio monitoring. Use an EVS 875 appraisal for: bank mortgage collateral, legal disputes, insurance, and official valuations.
2. Data sources
Estiq combines state registry data with the Estiq market database. The notarised sale price of an individual flat is not published in Estonia, so the estimate is built from the asking prices of similar nearby flats, which are brought to transaction level using Land Board sales statistics:
| Source | Data used | Update | Role in AVM |
|---|---|---|---|
| Maa- ja ruumiamet, tehingute andmebaas Maa- ja ruumiamet, tehingute andmebaas (tehingute andmebaas) |
Notarised sales statistics: median prices and transaction counts by area, construction period and month | Monthly | Measures the gap between asking and sale prices (0.95–0.97 in Tallinn), sets the regional price level outside Tallinn, and is the fallback where there are no nearby comparables. |
| Ehitisregister (EHR) National Building Register |
Building specifications: construction year, floor count, building type, total area, kasutusluba status | Weekly | Used to pre-fill the form and to determine the construction era and lift for the subject and the comparables. |
| MKM Energy Database Ministry of Climate |
Energy performance certificates — A through H class ratings for individual buildings | Weekly | Energy class coefficient applied as a valuation adjustment. A/B adds premium; F–H applies discount. |
| Uber H3 Spatial indexing |
Hexagonal geocell grid at resolution 9 (~0.1 km² per cell) covering all of Estonia | Daily | Regional price level outside Tallinn (resolution 9 → 8 → 7), maps and area statistics. In Tallinn, distance is measured in metres. |
| Statistikaamet Statistics Estonia |
Macro price indices: quarterly housing price index by county and property type | Weekly | Used to detect rapid market shifts and apply trend correction between data refresh cycles. |
| OpenStreetMap OSM Contributors |
Points of interest: schools, parks, transit stops, commercial zones, green space polygons | Weekly | Proximity signals fed into the Estiq Life Index. Not used in core price calculation. |
| Estonian Environmental Board Keskkonnaamet |
Air quality monitoring station data — AQI readings across Estonian cities | Daily | Feeds the Air Quality component of the Estiq Life Index™ displayed on district and address pages. |
An asking price is not a sale price, so comparable prices are brought to transaction level: according to the Land Board, flats in Tallinn and Tartu sell on average 3–5% below the final asking price. In Pärnu and smaller towns the gap is about 10%, and in Ida-Viru County 25–50% — which is also where the model's error is larger (see section 5).
3. AVM calculation — step by step
The steps below apply to Tallinn flats (model V9, in production since 22.09.2026). Elsewhere in Estonia the V7 model is used: it values a flat from the regional price level of H3 geocells (resolution 9 → 8 → 7) and shrinks small-sample estimates towards the wider area (empirical Bayes). A valuation takes a few seconds:
1. Address resolution and building matching
EHRThe entered address is normalised against the Ehitisregister (EHR), and building attributes — construction year, floor count, lift and building type — are loaded automatically. If the construction year is missing, it is inferred from data on other flats in the same building.
2. Comparable selection
24 months • 400 m → 3 km • at least 25 flatsIn Tallinn the model searches the Estiq market database for flats offered for sale around the subject over the last 24 months. The search radius starts at 400 m and widens (800 m, 1.5 km, 3 km) until at least 25 flats are found. Each flat counts once, at its latest known price. There are no hard size or age filters: differences are adjusted in the next steps.
3. Adjusting every comparable
Size • construction era • time (+2%/yr)Each comparable's price per m² is converted to the subject flat: size (small flats have a higher €/m²; the curve was measured on 5,871 Tallinn flats) and construction era are accounted for, and older price data is brought to today's prices (+2% per year).
4. Condition and energy class
New +4% • Renovated +6% • Fair −7% • Needs renovation −14% • A +4% … H −9%Condition factors are relative to a flat in good condition and are applied to both the subject and the comparables, so the difference between them cancels out. The energy certificate (EHR) adjustment is relative to class C and applied at half strength, because energy class is strongly tied to building age, which is already accounted for.
5. Weighting
Distance 350 m • half-life 9 months • same building 6×Closer, more recent and more similar (size and era) flats weigh more. Flats on the same street get 1.5× weight and flats in the same building 6×. The price per m² is the weighted median.
6. From asking to transaction price, and floor
× 0.97 • Ground floor (older building) −6% • 5th floor+ without lift −11%The weighted median is multiplied by 0.97, because Land Board sales statistics show Tallinn flats sell on average 3–5% below the final asking price. A floor adjustment follows: ground floor −6% in older buildings (+1% in new ones), 5th floor or higher without a lift −11%, and from the 6th floor in buildings with a lift +1.5% per floor (up to +12%).
7. Price range
80% range • ±8–35% • average ±19%The width of the range depends on how widely the comparables' prices spread and how many there are. It is calibrated so that the real price falls inside it four times out of five (see section 4).
4. Price range and confidence
Every estimate comes with a price range. In Tallinn it is an 80% range, calibrated so that the real price falls inside it four times out of five. In the blind test (May–September 2026) the price fell inside the range 78–81% of the time, and the range averaged ±19%. Its width depends on how many similar flats are nearby and how similar they are:
High confidence
Many similar flats nearby with consistent prices. The range is at its narrowest (from ±8%). Typical in standard apartment-block districts such as Mustamäe and Lasnamäe (median error 5.8% and 7.6%).
Medium confidence
Fewer similar flats, or their prices spread more. The range is wider. Consider checking the estimate with a broker before a transaction.
Low confidence
Thin data: an atypical flat, a sparse area or a new building without comparables. The range can reach ±35%. Use it as a starting point only.
The “effective number of comparables” accounts for weights: one flat in the same building counts for more than several flats further away. Outside Tallinn (model V7) the range is computed from the statistical spread of comparable prices (capped at ±30%); its coverage has not yet been tested separately.
5. Empirical Model Accuracy & Backtest Results (2026)
Estiq measures model accuracy with a blind out-of-sample backtest on data the model has never seen. Every test flat is valued with an as_of cutoff: the engine only sees data that existed on the test flat's valuation date. The flat's own earlier price records (same street, building and floor area) are also hidden — roughly 73% of recent Tallinn price records are repeat records of the same flat, and without hiding them a model could simply "read back" its previous price.
Reference price: the notarised transaction price of an individual flat is not published, so the test's "true price" is the flat's latest asking price × 0.97. In Tallinn, comparable flats sell on average at 0.95–0.97 × the final asking price. Model V9 was tuned on data from January–April 2026; the test window 17.05.–14.09.2026 is its first independent check.
Tallinn — Estiq AVM V9 (in production since 22.09.2026)
Random sample: 600 flats (one row per flat) out of 3,421 Tallinn flats, blind test. All models were scored on the same sample.
| Model (Tallinn, n = 600) | Median error | Bias | Within ±10% | Within ±20% | Price in range |
|---|---|---|---|---|---|
| V9 — adjusted comparables (k-NN), current | 9.5% | −1.7% | 53% | 83% | 78% |
| V6 — previous Tallinn model | 12.0% | −0.2% | 41% | 74% | 71% |
| V7 — calibrated micro-level model | 11.7% | +1.3% | 43% | 75% | 72% |
| V8 — calibrated | 12.6% | +1.9% | 39% | 70% | 69% |
A second independent random sample (600 flats) gave V9 a median error of 8.6% (V6: 11.0%), with 56% of estimates within ±10% and 87% within ±20%. A negative bias means V9 values slightly below asking price × 0.97.
Accuracy by building age and district (Tallinn)
The biggest gains are in new developments and flats with an unknown construction year: V9 adjusts every comparable to the subject flat's size, era, condition and energy class, and infers a missing construction year from data on other flats in the same building.
| Segment | n | V9 median error | V6 median error | V9 within ±10% |
|---|---|---|---|---|
| New build (2015+) | 180 | 8.6% | 12.8% | 57% |
| 2005–2014 | 52 | 10.7% | 13.7% | 46% |
| Before 2005 | 329 | 9.7% | 11.0% | 51% |
| Construction year unknown | 39 | 7.9% | 16.3% | 54% |
V9 median error by district (previous model V6 for comparison):
Districts with small samples (Nõmme, Pirita) carry more statistical uncertainty. Pirita is the only district where V9 did not beat the previous model.
Rest of Estonia (model V7)
V9 is used for Tallinn flats with a known location; elsewhere in Estonia flats are valued by the location-calibrated V7 (H3 geocell expansion res-9 → res-8 → res-7 and empirical Bayes shrinkage). The same blind test across Estonia: 1,183 flats in 46 towns and settlements (at most 40 flats per settlement) — median error 12.1%, bias +0.9%, 44% of estimates within ±10% and 68% within ±20%; the simple average of per-settlement median errors is 15.1%.
In settlements around Tallinn (e.g. Haabneeme 5.4%, Saue 5.3%, Saku 5.9%, Laagri 6.0%) the error is comparable to Tallinn. In Ida-Virumaa (Narva, Sillamäe, Kohtla-Järve) and small county towns the error is higher: there are few transactions and the gap between asking and transaction prices is much wider than in Tallinn, so a comparison based on asking prices is also less reliable there.
6. Estiq Life Index™
The Estiq Life Index is displayed on all district, neighbourhood, and address pages. It is a composite quality-of-living score calculated per 1 km² grid cell from four data streams:
🛡️ Crime Safety
Recorded incidents from Police and Border Guard Board (PPA) open data on the H3 grid. The 0–10 index is the incident rate per 1,000 residents on a log scale; with fewer than 10 incidents no level is shown (see 6.1).
💨 Air Quality (AQI)
Hourly PM10 and NO₂ classes from Keskkonnaagentuur (Estonian Environment Agency) data on the H3 grid (Tallinn, Tartu, Pärnu, Ida-Viru). Building pages show the 12-month class split; elsewhere air is not rated (see 6.1).
👥 Population Density
Residents per 1 km² from Statistics Estonia. High density signals urban amenities; used for liveability scoring and walkability proxies.
🏫 Schools & Green Zones
Catchment quality from school performance data (akadeemiline reiting) plus green space polygon coverage from OpenStreetMap.
6.1 Environment and safety on building pages
Each building page has an “Area at a glance” strip and an “Environment & safety” section. The indicators describe the building’s location on the H3 grid (Uber H3, resolution 9, ~0.1 km²), not a specific apartment. Where an address has no data, the page says so rather than showing a generic default.
| Indicator | Data | Calculation | Levels and limits |
|---|---|---|---|
| Incident level | Police and Border Guard Board (PPA) open data: recorded incidents over the last 24 months. | Incidents are weighted by type and divided by the cell’s residents (per 1,000). Index 0–10 = log(rate + 1) / 8.8 × 10. Building cell ~0.1 km²; if it has none, the ~0.7 km² or ~5 km² area. Change = last 12 months vs the 12 before. | Below 2.5 low · 2.5–5 moderate · 5–7 above average · above 7 high. With fewer than 10 incidents no level is shown, only the count. |
| Air | Estiq’s hourly air layer from Keskkonnaagentuur (Estonian Environment Agency) data on the H3 grid: Tallinn, Tartu, Pärnu, Muuga and the Ida-Viru towns. | The source gives PM10 and NO₂ as five classes, not exact concentrations. Each hour takes the worse of the two; we show the 12-month split (two best, middle, two worst classes). With under 24 readings in the cell, cells within ~600 m are pooled. Filler records and OpenWeatherMap fallback rows are excluded. | Share of poor hours: under 2% low · 2–8% moderate · 8–20% elevated · over 20% high. Outside the covered areas air is not rated. |
| Noise | Estiq noise model: OpenStreetMap road network with typical speed and traffic flow per road class. | Simplified CNOSSOS-EU approximation: each nearby road’s level by distance, summed to the day-evening-night indicator Lden (dB). | The WHO recommends below 53 dB Lden for road traffic. Rail, aircraft and industry are not modelled. |
| Radon in the ground | Estonian radon-risk map (Geological Survey of Estonia): radon in soil gas, kBq/m³. | The map classes are carried onto the H3 grid. Where a cell has no measurement the class is interpolated from the nearest ones, and the building page says so. | Under 30 low · 30–50 normal · 50–250 high · over 250 very high (kBq/m³). Above 50, new buildings need radon-safe construction (EVS 840). The indoor reference level of 300 Bq/m³ is a separate measure. |
| School and transit | OpenStreetMap and Estiq school data (state exam results). | Nearest place in a straight line; walking time = distance / 75 m per minute. | Places more than 2 km away are not shown as walking distance. |
These indicators do not change the AVM estimate. The air and noise levels feed the building score (“Air quality & environment”, weight 10%); the incident index feeds “Safety” (weight 10%).
7. Limitations and what the AVM cannot see
The model's median error is 9.5% in Tallinn and 12.1% elsewhere in Estonia (see section 5). Beyond that, it is important to know its boundaries:
- Physical condition: The model uses user-reported condition, not an on-site inspection. Unreported defects, structural issues, or undocumented renovation work are not captured.
- Legal encumbrances: Mortgages, easements, debt with the building management company (korteriühistu), and other legal constraints on the title are not reflected.
- Rental income and yield: The AVM produces a capital value estimate only. Rental yield analysis is not included.
- Renovation specifics: A full kitchen and bathroom renovation increases value beyond the "Excellent" condition coefficient, but specific renovation costs and outcomes are not modelled at the individual property level.
- Atypical properties: Heritage buildings, properties with significant plot value, commercial-residential mixed-use units, and properties with unusual configurations may be less accurately valued by the AVM.
8. Frequently asked questions
How accurate is Estiq's property valuation?
Estiq measures accuracy with a blind test on flats the model has never seen (17.05.–14.09.2026). In Tallinn the median error is 9.5%: 53% of estimates fall within ±10% and 83% within ±20%. Across the rest of Estonia (1,183 flats in 46 towns) the median error is 12.1% — Tartu 13.3%, Pärnu 14.8%, Narva 22.7%. Accuracy is highest for standard flats in active areas and lower for atypical properties and areas with few sales. This is a statistical estimate, not a certified EVS 875 appraisal.
What data sources does Estiq use for property valuation?
The estimate is built mainly from the Estiq market database: the prices of similar nearby flats. Land Board (Maa- ja ruumiamet) sales statistics bring them to transaction level and set the regional price level. Estiq also uses the Ehitisregister (EHR) for building data, the energy certificate database for A–H classes, Uber H3 geocells for area statistics, and Statistics Estonia, OpenStreetMap and Environmental Board data for market statistics and the Estiq Life Index.
What is the difference between Estiq's AVM and an EVS 875 certified appraisal?
Estiq's automated valuation model (AVM) is a statistical tool for instant market price estimation. It uses verified transaction data but does not include a physical inspection of the property. An EVS 875 certified appraisal is a formal document produced by a licensed valuer (hindamisekspert) after an on-site visit. Estonian banks require EVS 875 appraisals as mandatory collateral documentation for mortgage lending. Estiq's AVM is suitable for market monitoring, pricing decisions, and broker briefings — not for bank collateral.
What does the confidence interval mean in Estiq's valuation?
Every estimate comes with a price range. In Tallinn it is an 80% range: in the blind test the real price fell inside it 78–81% of the time, and it averaged ±19% (narrowest ±8%, widest ±35%). The range is narrower when many similar, consistently priced flats are nearby (high confidence) and wider for atypical flats or sparse areas (low confidence). Outside Tallinn the range is computed from the spread of comparable prices (capped at ±30%).
How often is Estiq's property data updated?
Estiq's transaction database is updated monthly from the Maa- ja ruumiamet, tehingute andmebaas (tehingute andmebaas). Building data from Ehitisregister and energy certificate data from MKM are synchronised on a weekly basis. District-level market statistics — including median price per m², transaction volume, and 3-month trend indicators — are recalculated monthly after the new data import. This ensures that all valuations and district pages reflect the most current available market data.
Data attribution:
Try the AVM on your property
Free for any Estonian address. No login required. Results in under 2 seconds.