What is an automated valuation model, and how insurers and banks use it

What an AVM is, and what it is not

An automated valuation model is a statistical model that turns a description of a property (where it is, how big it is, what type and condition it is in) into an estimate of its market value, with no inspection and no human judgment applied to the individual case. The European AVM Alliance (EAA), whose standards took effect on 1 March 2022 in their third edition, places AVMs at the top of a ladder of four statistical valuation methods: single-parameter valuations (a price per square meter for an area), house price indices (a change applied to a previous value), hedonic models (value as a function of attributes) and comparables-based AVMs, which select and adjust similar recent sales for each subject. What makes the output an AVM is that it is property-specific, reproducible and produced without a valuer.

An AVM is not an appraisal. An appraisal is an opinion of value by a qualified person who takes responsibility for it; an AVM is a measurement with a known error distribution. It is also not an index: an index says how prices moved, an AVM says what one property is worth now. And it is not a rebuild cost. The market value an AVM produces includes land, location and demand, while the sum insured on a property policy is usually a construction cost that excludes all three. One 2025 expert appraisal of a 70 m2 panel flat illustrates the gap: rebuild value 1,400 EUR/m2, official collateral value 1,064 EUR/m2, and asking prices for comparable flats between 1,667 and 2,190 EUR/m2. The rebuild cost guide covers the insurance side.

Inputs: what the model needs to know

A residential AVM reads a short list of facts about the property and a long list of facts derived from its coordinates. The table lists the inputs that move the number most, where they usually come from, and what goes wrong when they are missing or wrong. The failure modes are real cases from one production model, not hypotheticals.

Inputs and failure cases from the Bytero Engine model documentation.

Model families: hedonic regression versus gradient boosting

Two families dominate. A hedonic model writes log price as a weighted sum of attributes, so every coefficient can be read and defended. Gradient-boosted trees (XGBoost and its relatives) learn interactions and non-linearities automatically and are the usual production choice where accuracy is the goal. The benchmark below fitted both on the same 81,311 apartment sale records, the same 80/20 split (64,148 training rows, 16,037 test rows) and the same log-price target, so the difference is the model form and nothing else.

Bytero benchmark of a linear and a gradient-boosting apartment sale model on the same Slovak hold-out, 2026; MAPE = mean absolute percentage error, APE = absolute percentage error.

Error metrics: how AVM accuracy is measured

The EAA standards split accuracy into bias and dispersion. Bias is the systematic tendency to over- or under-value, measured by the average or, preferably, the median error against a benchmark value such as a sale price. Dispersion is the spread of errors regardless of sign, and the standards name three ways to report it: the standard deviation, the average absolute error, and the percentages of valuations within 10, 15 and 20 percent of the benchmark. The 20 percent bucket is the one most often quoted, on the reasoning that both the model and the benchmark carry about 10 percent of inherent uncertainty, so two values can differ by 20 percent without either being wrong. Like-for-like comparisons between vendors must use the same bucket on the same sample.

Two metrics from the benchmark above deserve translation. Median absolute percentage error (MedAPE) is the error of the typical property: half of test properties were valued within 7.6 percent by the boosted model and within 14.5 percent by the hedonic one. Mean absolute percentage error (MAPE) is dragged up by the tail of hard cases, which is why it reads 12.9 and 23 percent for the same models. R2 on the log scale is the share of price variation explained; R2 on the euro scale is unstable for log models, dominated by a few luxury properties, and should not rank models. Error must be measured on a hold-out set the model never saw, and for spatial data the split should keep neighboring properties on the same side, or leakage inflates every metric.

Two further quantities are often confused with accuracy. The EAA hit rate is the share of properties that meet the model's input requirements and still receive a value; a model can be accurate and refuse thin cases. A confidence measure is a per-property estimate of reliability; the standards require it to be predictive, meaning translatable into a forecast standard deviation, and EBA paragraph 210 requires one whenever a model supports a desktop valuation. One production model computes a 0 to 100 confidence score from weighted completeness of the 35 input columns it reads, with badges below 85 and below 60.

What the EBA and the CRR allow

The rules for European banks sit in two places: the Capital Requirements Regulation (CRR, Regulation (EU) No 575/2013) and the EBA Guidelines on loan origination and monitoring (EBA/GL/2020/06), which apply from 30 June 2021. The guidelines call an AVM an advanced statistical model and are explicit about when it may support a valuation and when it may stand alone.

EBA/GL/2020/06 final report, section 7; Regulation (EU) No 575/2013, Article 208.

How insurers use an AVM

Property insurers use market value for a different purpose than banks. The sum insured on a building is normally a rebuild cost, but for flats several carriers write their terms on market value, sometimes only where it exceeds the rebuild figure, and pay a condemned flat at market value up to a cap. An AVM is therefore the reference for a sum-insured check on flats: does the sum on the policy reach the value the terms describe? Where the policy basis is rebuild cost, the AVM is the wrong number and a cost engine is the right one; a portfolio check needs both, applied per carrier and per property type.

The second insurer use is renewal and indexation. Sums insured are indexed annually by a construction or residential price index, and a portfolio-wide AVM run shows where indexation on flats has fallen behind the market or run ahead of it. The third is underwriting context: market value beside rebuild cost identifies the properties where a total loss would leave the owner unable to buy an equivalent, which is where disputes start. A rebuild cost derived from a building's LiDAR volume closes the loop; see the LiDAR rebuild-cost guide.

How banks use an AVM

The core bank use is loan-to-value monitoring. CRR Article 208(3) requires the value of residential collateral to be monitored at least every three years and commercial collateral at least yearly, and allows statistical methods for that monitoring and for identifying which properties need a revaluation. An AVM run across the mortgage book gives a current value and LTV per loan and a list of loans that crossed a trigger; EBA paragraph 223 sets the quality bar: granular enough for the asset and built on enough observed transactions and appraisals.

Revaluation is where the guideline draws its line. After a material decline, paragraph 225 requires a valuer, who may lean on a model but not hand the job to it; absent such a trigger, paragraph 226 allows the model alone if it meets Section 7.4. Origination stays with the valuer except for the paragraph 210 desktop derogation. Banks that adopt an AVM therefore keep three artifacts for their supervisor: the model documentation, the back-test against observed prices, and the per-valuation audit trail described next.

What an audit trail contains

A valuation that cannot be reproduced is not evidence. EBA paragraph 238 requires institutions to understand their models' methodology, inputs and assumptions and keep documentation current; paragraph 237(b) requires back-testing against observed transaction prices. Per valuation, the record should hold enough to recompute the number a year later.

• Model identity and version. The model family, the trained artifact and its version label (one production model stamps every value with a label such as v3.3 and routes houses and cottages to separately labeled models), so a retrain never silently changes an old figure.
• Input snapshot. Every attribute as submitted and as enriched (coordinates, distances, postcode density, condition score), with a machine-readable reason for anything missing (no_gps, no_area).
• Comparables or cell evidence. The selected sales and adjustments for a comparables model; the location cell and its sample size for a hedonic or boosted one.
• Post-model adjustments. Any asking-to-sale correction with its parameters: one engine applies a regional baseline of about 6 to 9 percent, a rate-regime factor and a staleness factor rising from 1.0 for a fresh offer to 2.0 for one a year old.
• Confidence and range. The confidence score, a lower band (a quantile model emits a 25th percentile beside the median) and the forecast standard deviation the confidence maps to.
• Data vintage and timestamp. The dates of the training data and of the market data read at prediction time; saved valuations stay frozen unless a price-relevant input changes.
• Back-test reference. The hold-out report the model passed before release: R2, MAPE, MedAPE and the share within 10 and 20 percent, per segment.

Limits

An AVM is only as good as the prices it learned from. Where sale prices are private, models train on asking prices and must correct for the gap to sale, which is neither constant nor small: one calibration puts the regional baseline between roughly 6 percent in a tight capital-city market and 9 percent in rural areas, widening when mortgage rates rise and as an offer ages unsold. Reported without that correction, the model measures what sellers hope for.

Accuracy is uneven by segment. Flats in dense markets are the easy case; houses (about 13 percent median error in one production model) and cottages (15 to 18 percent) are harder because plot size, remoteness and condition matter more and are recorded less. Unique properties, land and anything outside the training distribution should return low confidence rather than a confident wrong number. Interpretability has a price too: on the benchmark above, the readable hedonic model paid about 10 points of MAPE and double the median error for its transparency.

Finally, an AVM answers one question. It does not give the rebuild cost an insurer needs or the depreciated technical value a court expert reports; those are separate, deterministic calculations. Bytero runs a gradient-boosted model of the kind benchmarked here, versioned per valuation with a confidence score and an asking-to-sale correction, alongside a separate rebuild-cost engine; the method is described on the Bytero Engine page and availability by market on the coverage page.

Questions