How AVM accuracy is measured

What the rules ask for

CRR Article 208(3a), added by CRR3, lets a bank monitor collateral with a model that is "valid and accurate, and subject to robust and regular back-testing against the actual observed transaction prices", built on "a sufficiently large and representative sample, based on observed transaction prices". Paragraph 237 of the EBA Guidelines on loan origination and monitoring uses the same words. Paragraph 210 adds that a model supporting a desktop valuation must carry "a confidence measure to indicate the robustness of the value proposal".

The rules say what to test against. They do not say which numbers to report. For that, lenders and vendors in Europe mostly use the European Standards for Statistical Valuation Methods of the European AVM Alliance, now in their third edition.

The accuracy measures

The Alliance defines the error of one valuation as the result minus the benchmark value, divided by the benchmark value. Bias and dispersion are both built from that error.

The 20 percent margin is the one quoted most often, because the standards allow that the model and the benchmark can each be about 10 percent off (section 5.2).

Hit rate is the term people misread most. In the standards it measures how often the model returns a value at all, and says nothing about how close the values are. A model can have a 99 percent hit rate and still miss by 30 percent on half its answers.

Mean absolute percentage error (MAPE) is not in the table because the standards do not lead with it. It is useful, but a few very wrong values pull it up, so it belongs next to the dispersion figures and a histogram.

What counts as a benchmark

Section 4.6 of the standards says that "A Benchmark Value consists of either a confirmed Sale Price or Surveyor Valuation", and "Asking Prices or estimated values that do not qualify as Surveyor Valuations are not considered as valid Benchmark Values."

The standards also describe two ways to run a test:

• Lender test. The bank picks recent cases, sends them to the vendor, and reveals the sale prices or valuations only after the results come back. It is blind by design. Samples tend to be small.
• Bulk test. The vendor draws cases and benchmarks from its own database and keeps them out of the model while it values them. Samples can be large, but the bank has to take the blindness on trust.

Either way the test must be out of sample, and "the results must not be tested in circumstances more favourable than those to be expected in real-life usage."

How a test can flatter a model

1. 1

   The same property, twice

   A flat offered for sale twice, on two websites or again at a new price, is one property. If one copy trains the model and the other tests it, the model is graded on a flat it has already seen. Duplicates should be merged before the test, and any copy of a test property that remains should be kept out of its comparables.

2. 2

   A friendly sample

   A model does best where it has many close neighbors. A harder test removes all properties near the test property from the training data and shows how the model does where evidence is thin.

3. 3

   Inputs the bank will not have

   A model tested with the year of construction the seller states can look better than it will be when it gets the year from a register. The test should use only the inputs the bank holds at monitoring time.

4. 4

   Grouping by the benchmark

   Splitting results by the true price shows cheap properties overvalued and expensive ones undervalued, partly as a statistical artifact, because the most expensive properties in any sample include those whose price was unusually high. Group by something the model knows, such as floor area, to read size effects.

Why a test against asking prices is not enough

Asking prices usually sit above sale prices. 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. A model that matches asking prices will overstate sale prices by a similar margin unless it corrects for that gap, and the correction itself needs sale prices to be checked.

A test on properties offered for sale also says little about a lender's own book, because mortgaged homes are a different mix from homes on the market. Our page on [asking and sale prices in Slovakia, Czechia, Poland and Hungary](/guides/asking-vs-transaction-prices-cee) lists the sale-price data each market publishes.

How a bank can test an AVM vendor

1. 1

   Take a few hundred recent completed sales or full valuations from your own files that the vendor cannot have seen.

2. 2

   Send the address and the inputs you would have at monitoring time, nothing more.

3. 3

   Ask for a value, a confidence measure and the model version for each case.

4. 4

   Compare: median error, the share within 10, 15 and 20 percent, a 5 percent histogram, the same by region and size, and the error spread per confidence grade.

5. 5

   Repeat each year, and after each model change.