Valuation

What an automated valuation model is

An automated valuation model (AVM) is a statistical model that estimates the market value of one property from its location and recorded characteristics. Banks use AVMs to monitor the value of mortgage collateral, which CRR Article 208(3a) has allowed under six conditions since 1 January 2025, and insurers use them to check sums insured on flats. Under the EBA guidelines on loan origination a model may support a valuer when a loan is made, and the valuer stays responsible for the value.

What an AVM is

The European AVM Alliance, whose standards have applied in their third edition since 1 March 2022, groups statistical valuation methods into four types. House price indices measure how prices in an area moved. Single parameter valuations apply one figure, usually a price per square meter, to a property. Hedonic models describe value as a function of the property's attributes and its location. Comparables-based AVMs select a set of similar properties for each property they value and calculate the value from them. In the strict sense of the standards, an AVM is the last type; in everyday use the term covers any model that values one property at a time.

An appraisal is an opinion of value by a qualified valuer who signs it and answers for it. An AVM calculates a value from data by a fixed method, so the same inputs give the same result, and its error can be measured on properties whose price is known. An index answers a different question, how prices moved, and a rebuild cost answers another one, what it costs to construct the building again. The rebuild cost guide covers the insurance side.

What a residential AVM reads

A residential AVM reads a short list of facts about the property and derives much more from its location. The inputs below move the value most.

Input	Why it matters	Usual source
Address or coordinates	Location sets the price level, and the model learns that level for each area from the data	Address register, cadastre, geocoder
Floor area	Value rises with size, and the price per square meter usually falls as flats get larger	Cadastre, the loan file
Property type	Flats, houses and land trade on different scales, so many vendors run a separate model for each	Cadastre, the loan file
Year of construction	Age changes value through building type, condition and energy standard	Cadastre, building register
Floor and number of storeys	Ground and top floors often trade at a discount or a premium, depending on the building	Cadastre, the loan file
Condition	A renovated flat and an original one in the same building can differ widely, but condition is rarely recorded in a register	Valuer's report, the borrower

A model that needs an input the bank does not hold at monitoring time will be less accurate in use than in its own test.

Model families

Three families of model are common. Each can be tested the same way, on properties whose sale price is known and that the model never saw.

Family	How it prices	Strengths	Weaknesses
Hedonic regression	The log of the price as a weighted sum of attributes	Each coefficient is a readable price effect; cheap to fit and to explain	A linear form misses interactions, for example between age and location
Gradient-boosted trees	The sum of many small decision trees, often with rules that keep value rising with area	Learns interactions and non-linear effects from the data	Needs feature attribution to explain a single value, and disciplined retraining
Comparables-based	Selects similar recent prices near the property and adjusts them to it	Close to how a valuer works; each value comes with its evidence	Needs dense, recent price data and gets weaker where there are few comparables

How AVM accuracy is measured

The European AVM Alliance standards split accuracy into bias and dispersion. Bias is the tendency to overvalue or undervalue, and the standards prefer the median error to the average. Dispersion is the spread of errors, reported most often as the share of values within 10, 15 and 20 percent of a benchmark value. The 20 percent margin is quoted most often, because the standards allow that the model and the benchmark can each be about 10 percent off.

The benchmark must be a confirmed sale price or a surveyor's valuation. The standards do not accept asking prices as benchmark values. The hit rate measures how often the model returns a value at all, and a confidence measure must be translatable into a forecast standard deviation. Our page on how AVM accuracy is measured sets out each measure and how a bank can test a vendor.

What CRR3 and the EBA guidelines allow

The rules for European banks sit in the Capital Requirements Regulation, as amended by CRR3 (Regulation (EU) 2024/1623), and in the EBA Guidelines on loan origination and monitoring (EBA/GL/2020/06), which have applied since 30 June 2021. The guidelines speak of advanced statistical models; neither text uses the term AVM.

Stage	What the rule says	Where
Policy	The bank sets out the valuation approaches a valuer uses and the use of advanced statistical models for each type of collateral	EBA paragraph 206
Origination	An internal or external valuer assesses the property with a full visit, inside and outside	EBA paragraph 209
Origination, exception	For residential property in well-developed and mature markets, a valuer may do a desktop valuation supported by a model that carries a confidence measure; the valuer remains responsible	EBA paragraph 210
Monitoring	Values are monitored at least once a year for commercial and once every three years for residential property, and a model may do the monitoring and pick the properties to revalue if six conditions are met	CRR Article 208(3)(a) and (3a)
Review after a possible material decline	A qualified valuer independent of the credit decision reviews the value; a model may support the valuer but may not be the sole means	CRR Article 208(3)(b); EBA paragraph 225
Other revaluations	A valuer or an appropriate statistical model may update the value	EBA paragraph 226
Model criteria	Property and location specific, back-tested against observed transaction prices, built on a large representative sample of transactions and on up-to-date data of high quality	CRR Article 208(3a)(b); EBA paragraph 237

EBA paragraph 226 was written against the old Article 229. Whether a model-only update meets the independent valuer test in the new Article 229(1) is not settled; our page on whether a bank can use an AVM under CRR3 goes through each condition.

How banks use an AVM

The main use is monitoring. A model run across the mortgage book gives a current value and loan-to-value ratio for each loan and a list of loans that crossed a trigger the bank has set. Since CRR3, those monitoring values also feed the six- or eight-year average that caps any upward revaluation of a property value (Article 229(1)(e)), so the quality of the model now reaches the capital figure. Our page on CRR3 property value works the cap through on one flat.

A second use runs the other way. EBA paragraph 233 asks banks to assess the accuracy of their valuers, for example by back-testing valuations with advanced statistical models. A bank that adopts a model keeps the model documentation, the back-test against observed prices and a record of each valuation for its supervisor.

How insurers use an AVM

The sum insured on a building is usually a rebuild cost, and for that a cost calculation is the right tool. For flats, some policies set the sum insured or the payout by market value, and there an AVM is the reference for checking whether the sum on the policy matches what the terms describe.

Insurers also index sums insured each year, and a market value run across a portfolio of flats shows where indexation has fallen behind prices or run ahead of them. A rebuild cost derived from a building's measured volume covers the houses; see the LiDAR rebuild cost guide.

What a valuation record should hold

EBA paragraph 238 asks banks to understand the methodology, input data and assumptions of their models and to keep the documentation up to date. For each valuation, the record should hold enough to calculate the same number a year later.

Model and version. The model family and the version that produced the value, so that a retrained model does not change an old figure without trace.
Inputs. Each input as received and as derived, and a reason for each one that was missing.
Evidence. The comparables and their adjustments for a comparables model, or the area and sample behind a hedonic or tree model.
Adjustments. Any correction applied after the model, such as a conversion from asking to sale price, with its parameters.
Confidence and range. The confidence measure, the range around the value and the forecast standard deviation it maps to.
Dates. The date of the valuation and the dates of the data the model read.

Limits

A model learns from the prices it is given. Where sale prices are not published property by property, models learn from asking prices and must correct for the gap to the sale price. That gap varies by place and over time. In Poland, where the central bank publishes both, the mean sale price per square meter on the secondary market was 84 to 99 percent of the mean asking price across 17 cities in the second quarter of 2026.

Accuracy also differs by segment. Flats in dense markets are the easier case. Houses are harder, because plot size, condition and fittings matter more and are recorded less, and unusual properties and land should come back with low confidence. An AVM gives a market value; the rebuild cost an insurer needs and the technical value a court expert reports are separate calculations. Where Bytero Engine is available is on the coverage page.

Questions

Is an AVM the same as an appraisal?

No. An appraisal is a valuer's opinion of value, signed by the valuer. An AVM calculates a value from data by a fixed method. Under EBA paragraph 210, a valuer who uses a model for a desktop valuation remains responsible for the value.

Can a bank use an AVM when it grants a loan?

Only to support a valuer. Paragraph 209 of the EBA guidelines asks for a valuer with a full visit; paragraph 210 allows a desktop valuation by a valuer supported by a model for residential property in mature markets, where the model carries a confidence measure.

Can a bank use an AVM to monitor collateral?

Yes. Since 1 January 2025, CRR Article 208(3a) allows models to monitor collateral values and to pick the properties that need a revaluation, under six conditions that include back-testing against observed transaction prices and independent validation.

What accuracy should an AVM have?

No regulation sets a threshold. The European AVM Alliance standards measure bias by the median error and dispersion by the share of values within 10, 15 and 20 percent of a sale price or a surveyor's valuation.