Method comparison

Automated valuation vs manual appraisal: when each is allowed

Under the EBA guidelines on loan origination and monitoring, a statistical model cannot value collateral on its own at origination or after a material decline. It can support a valuer there, and it can stand alone for routine monitoring and for revaluations that no decline has triggered. Everything else on this page is the working: cost, turnaround, consistency, auditability, what each method gets wrong, and the hybrid workflow that satisfies the rule.

The rule, paragraph by paragraph

EBA/GL/2020/06 has applied since 30 June 2021. Section 7 covers valuation, monitoring and revaluation of immovable property collateral; the paragraph numbers below are from the final report.

Origination (paragraph 209)The value of all immovable property collateral is assessed by an internal or external valuer using a full visit with internal and external assessment of the property.

Origination, residential derogation (paragraph 210)For residential real estate in well-developed and mature property markets, the value may be assessed by a desktop valuation carried out by a valuer and supported by advanced statistical models. The models must include a confidence measure, the valuer remains responsible and must understand all inputs and assumptions, and if the confidence measure indicates low robustness the valuer should choose a method other than desktop valuation.

Monitoring (paragraphs 221 to 223)Institutions set the approach and the frequency of monitoring, with higher frequency for property in development, high LTV or lower credit quality. Any indices and statistical models used to monitor value must be sufficiently granular, appropriate for the type of asset and lending product, and based on a sufficient time series of observed transactions and appraisals.

Revaluation after a material decline (paragraph 225)When the review conditions in Article 208(3)(b) of Regulation (EU) No 575/2013 are met, the revaluation is carried out by a valuer who is potentially supported by appropriate advanced statistical models. Institutions should not use these models as the sole means of the revaluation.

Revaluation without a trigger (paragraph 226)When those review conditions are not met, the value may be updated either by a revaluation carried out by a valuer or by appropriate statistical models that account for the individual characteristics of the property and geographical area.

Model criteria (paragraphs 236 to 238)Models must be property and location specific at a sufficient granularity (postcode is the example given), valid and accurate with robust and regular back-testing against observed transaction prices, based on a sufficiently large and representative sample of observed transactions, and on up-to-date data of high quality. The institution is responsible for the model; the valuer remains responsible for the valuation made with it.

Czechia and SlovakiaThe EBA compliance table for these guidelines, last updated on 22 July 2025, records the Czech National Bank as intending to comply by 30 June 2021 and Národná banka Slovenska as complying, with the guidelines implemented into its supervisory procedures. A lender supervised by either applies Section 7 as written.

The guidelines do not use the term AVM. They speak of "advanced statistical models" and of "indices and statistical models", and the European AVM Alliance standards map their own definitions onto those terms. Read "advanced statistical model" as the regulated name for an automated valuation.

Side by side

Dimension	Automated valuation	Manual appraisal	Hybrid: desktop valuation by a valuer, model-supported
Cost per property	A fraction of a visit. The EBA impact assessment describes model valuation as quick and cost-efficient.	A professional's time plus travel and a signed report; the largest line in a collateral budget.	Model cost plus valuer review time; no travel, no appointment.
Turnaround	Seconds per address, a portfolio overnight.	Days to weeks: appointment, visit, report.	Hours to days; nothing to schedule.
Consistency	The same inputs give the same value every time. Drift is systematic, so it can be measured and corrected.	Varies by valuer and by day. Two valuers can reasonably differ on one property.	The model anchors the number; the valuer adjusts it with stated reasons.
Auditability	Inputs, comparables and a confidence measure logged per call; back-tests against transactions on file.	A signed report. The reasoning lives in the report and the valuer's file.	Both: the model log and a signed report that references it.
Origination	Not as the sole basis (paragraphs 209 and 210).	Required: a valuer with a full visit (paragraph 209).	Permitted for residential property in mature markets, valuer responsible (paragraph 210).
Routine monitoring	Permitted if granular, appropriate and built on observed transactions (paragraphs 221 to 223).	Permitted; rarely affordable at portfolio scale.	The usual design: model on the book, valuer on the exceptions.
Revaluation after a material decline	Not as the sole means (paragraph 225).	Required; the model may support the valuer.	Permitted: a valuer supported by the model (paragraph 225).
Revaluation without a trigger	Permitted (paragraph 226).	Permitted (paragraph 226).	Permitted.
Insurer sum-insured checks	Outside the EBA rule. Suited to screening a whole book for underinsurance on a rebuild-cost basis.	Used for large, unusual or disputed risks.	Screen with the model, refer the outliers.
Error profile	Blind to condition, interior, unpermitted works and anything not in the data. Errors cluster where comparables are thin.	Sees the building. Errors come from judgment and time pressure, and are hard to detect without a second opinion.	The model catches inconsistency; the valuer catches what the data cannot see.
Best used for	Portfolios, monitoring, screening, a second opinion on a report.	Origination, disputes, non-standard property.	Residential origination in mature markets; triggered revaluations.

"Permitted" and "not as the sole basis" follow the paragraph numbers in the section above. Cost and turnaround are described in relative terms on purpose: panel fees and model prices vary by market and by contract, and none are published here.

A worked cost example: monitoring 10,000 collateral properties

The ranges below are deliberately wide placeholders, not quotes. Put your own panel fees and model price in. Three designs, one year, 10,000 residential properties, with V the fee per physical appraisal, M the cost per model valuation, and E the share of the book escalated to a valuer.

Design	Formula	With V = 150 to 400, M = 1 to 10, E = 5 to 10 %	Regulatory position
Physical revaluation of every property	10,000 × V	1.5 to 4.0 million	Exceeds what the guidelines require for monitoring. Nobody runs this.
Model-only monitoring	10,000 × M	10,000 to 100,000	Permitted for monitoring (paragraphs 221 to 223) and for untriggered revaluations (paragraph 226). Not for a revaluation after a material decline.
Hybrid: model on all, valuer on the escalated share	10,000 × M + E × 10,000 × V	85,000 to 500,000	Satisfies paragraphs 225 and 226 if the escalation trigger includes the Article 208(3)(b) review conditions and the model’s low-confidence flag.

The point of the arithmetic is its shape, not its numbers. The hybrid's cost is dominated by the escalated share, not by the model price, so the design of the escalation trigger is the budget decision. The guidelines' own impact assessment says the same thing from the other direction: a broader use of advanced statistical models lets an institution "benefit from a quick and cost-efficient valuation", while a strict ban at origination "may increase the cost for consumers".

What each method gets wrong

The model cannot see the building

The European AVM Alliance defines an automated valuation as one obtained without conducting a physical inspection of the subject property. Everything the model does not have in its data, it cannot price: condition, the interior, unpermitted works, a view, a neighbor. Its errors cluster where comparable evidence is thin, which is why the guidelines require property- and location-specific granularity, back-testing against observed transaction prices, a representative sample and up-to-date data (paragraph 237), and why the residential derogation at origination hinges on a confidence measure (paragraph 210). The Alliance standards ask comparables-based models to report a confidence level or a forecast standard deviation with each result, and to demonstrate its reliability through lender and bulk tests.

The valuer cannot see the portfolio

A physical appraisal sees the building and is the only method the guidelines accept alone at origination. Its weaknesses are the mirror image of the model's: it is one professional's judgment on one day, two valuers can reasonably differ on the same property, and it does not scale to a book of ten thousand at any monitoring frequency a bank would pay for. RICS notes that speed, cost and the removal of human error are the factors driving adoption of automated methods, and that in practice the market is a spectrum of hybrids rather than a choice between a computer and a person. The guidelines agree: paragraph 238 leaves the institution responsible for the model and the valuer responsible for the valuation, in the same sentence.

Insurers: a different question, no EBA rule

An insurer checking a sum insured is not valuing collateral and is outside the loan origination guidelines altogether. The figure it needs is a rebuild cost, not a market value, and the failure it is screening for is underinsurance across a whole book rather than one wrong number on one policy. That is the case an automated method is well suited to: run every risk, flag the ones whose sum insured sits far from the modeled rebuild cost, and refer only those to a surveyor or a valuer. Bytero's rebuild cost is built for that screening step, from measured building geometry rather than a flat rate per square meter; which markets are live is on the coverage page, and nowhere else on this site.

A hybrid workflow that satisfies the guidelines

1. Model the whole bookRun the model at the monitoring frequency the policy sets. Paragraph 222 ties frequency to development status, carrying amount, LTV and credit quality, so the frequency is per segment, not per book.

2. Escalate on three triggersIndicators of a material decline in value (the Article 208(3)(b) review conditions), a low confidence measure on the model result, and property-specific information that casts doubt on the value proposal. Paragraph 224 asks institutions to write these triggers down.

3. Valuer on the escalated setA desktop, drive-by or full valuation by an internal or external valuer from the panel, with the model as a supporting tool and never as the sole means (paragraph 225). The valuer records why the model result was or was not adopted.

4. Back-test and documentTest the model regularly against observed transaction prices and keep the documentation current (paragraphs 237 and 238). The Alliance standards ask for lender tests and bulk tests of the confidence measure for the same reason.

5. Keep the panel and its rotationMaintain the panel of accepted external valuers with the relevant sector expertise (paragraph 211) and apply the rotation rule in Section 7.3, so the same valuer does not revalue the same property indefinitely.

Which to use

Use the model when

You are monitoring a book rather than deciding a loan: paragraphs 221 to 223 permit indices and statistical models for monitoring if they are granular, appropriate to the asset and built on observed transactions.
A revaluation is due but nothing in Article 208(3)(b) of the CRR has been triggered: paragraph 226 allows either a valuer or an appropriate model.
You want a second opinion on a valuer's report: the Alliance notes that lenders use comparables-based models to critically review valuations they receive, which is what paragraph 214 asks institutions to do.
The job is insurance screening, where no valuation rule applies and the cost of visiting every risk would exceed the premium.

Use a valuer when

You are originating a loan: paragraph 209 requires an internal or external valuer with a full visit, and paragraph 210 only relaxes that to a valuer's desktop valuation, model-supported, for residential property in mature markets.
Information indicates a material decline in value: paragraph 225 requires a revaluation by a valuer and says models should not be the sole means.
The model's confidence measure indicates low robustness, or property-specific information casts doubt on the value: paragraph 210 tells the valuer to choose another method.
The property is non-standard, in dispute, or the building itself is the question. A model that has never seen the building cannot answer for it.

How this page is maintained

Regulatory statements quote or paraphrase the EBA final report and the EBA compliance table, with paragraph numbers, and the method statements come from the European AVM Alliance standards and from RICS. No prices are published here because none of the public sources we rely on publishes them; the worked example is labeled as placeholder ranges for that reason. Corrections are welcome and will be applied.

Sources

Suggest a correction