How accurate is WoonWaarde?

We tested our model on 770 real home sales from the Kadaster (the Dutch land registry). In half of the cases the estimate was less than 7.2% off, and in 62% it was within 10%. If we also count the 176 sharply differing sales that we did not assess, the median error is 9.9%. Below is how we measured that, where the model works well and where it doesn't.

The result in four numbers

  • 7.2%median error between estimate and sale price
  • 62%of estimates within 10% of the sale price
  • 88%within 20% of the sale price
  • 770Kadaster sales tested, from 535 different homes

Test run on 19 September 2026 with the model as it runs now, without the human check that every report then gets.

How we tested

For each sale from the Kadaster, we acted as if it had yet to happen. The model was given only the home's features (floor area, year built, energy label, type) and the sales in the same street from before that date, up to three years back. We then put the estimate next to the actual purchase price.

The error is the difference between the estimate and the purchase price, as a percentage of the purchase price. The median error is the middle number: half of the estimates were closer, half further off. An average would be pulled up by a few outliers; here it is 10.3%.

What the test does differently from a report

  • Comparable sales. The test only uses earlier sales from the same street, up to three years old. In a report, a member of staff also chooses them from the same building or close by. So it isn't exactly the same.
  • Condition of the home. This is missing from the test, and in a report you fill it in yourself.
  • Review. The test runs without a person, a report does not. We have not yet measured whether that review makes the estimate better.
  • Size. The 770 sales come from 19 streets. That is a small base.

Which sales did not count

The file contained 1,214 Kadaster sales from 746 addresses. We could fairly test 770 of them. The rest dropped out, and that is shown below, so nobody thinks we only chose the easy cases.

ReasonSales
Tested770
No earlier sales in the same street within three years186
Purchase price more than 30% away from the street (for example a family sale or auction)176
Living area unknown71
Same-day resale (ABC transaction)6
Energy label unknown5

What the 176 differing sales do to the result

We excluded sales that differed by more than 30% from the street's price level, because those are often family sales or auctions. But that limit uses the actual purchase price, which you don't know before a home is sold. So we also show the result when they do count. Some of these sales are real market sales. The truth therefore lies between these two lines.

GroupSalesMedian errorWithin 10%Within 20%
Assessed7707.2%62%88%
With the 176 included9469.9%51%73%
From 2024, with the differing sales included1689.1%57%83%

Sales where a parking space or storage unit was in the same deed were counted. The model first takes those out of the purchase price. That applies to this test. In a report, such a sale has a footnote, and a member of staff checks whether it is comparable.

How confident the model itself is

Every report gets a confidence level: high, medium or low. It depends on how many comparable sales there are and how well they agree. Across the whole test, the estimate was off less often at high confidence, but the label is no guarantee.

ConfidenceSalesMedian errorWithin 10%Old range (± 8 to 12%)New range
High1505.4%79%73%± 11%
Medium3796.7%66%59%± 14%
Low24112.3%43%43%± 24%
All7707.2%62%56%

We first used ± 8%, ± 10% or ± 12%, depending on the number of sales. Only 56% of purchase prices fell within that. So we now measure the range: the new widths are chosen so that about 8 out of 10 assessed sales fall within them. We checked that on streets that had not been used to set the width: 81%.

Read this as a rough measure, not an exact probability. The base is small (19 streets), the 176 differing sales don't count here, and a member of staff can't remove all errors. From 2024, at high confidence 53% fell within the old ± 8% (36 sales), so high confidence can be wrong too.

Second test: 30 sales with assessed condition

The large test has no information on the home's condition. So we also tested 30 sales where someone assessed the condition from the listing. The estimate was fixed before we put the purchase price from the Kadaster next to it.

GroupSalesMedian errorWithin 10%Within 20%Within the range
All306.3%70%87%87%
High confidence115.0%100%100%100%
Medium confidence127.7%58%83%75%
Low confidence716.3%43%71%86%

When we weighed the estimate and the asking price together into an expected sale price, the median error was 3.9%, with 90% within 10%. The asking price alone was 5.6% off.

This does not replace the large test. 30 sales is few, especially per confidence level. The comparable sales here were mainly asking prices and not purchase prices, and nobody checked the final amount as happens with a report. For 24 of the 30 sales we already knew the purchase prices when we changed some of the calculation rules, so this test is not fully independent. Here we used the new ranges of ± 11%, ± 14% and ± 24%.

By city

CitySalesMedian errorWithin 10%
Rotterdam4066.4%65%
Den Haag2007.1%63%
Utrecht987.9%59%
Other cities (5, including Amsterdam)66too few per city to list separately

For Amsterdam the test is still too small: 17 sales, all in one street with very different pre-war buildings. There the model was clearly off, with a median error of 22%. That figure says little about Amsterdam as a whole, but we do mention it. We are expanding the test there.

By price range

Purchase priceSalesMedian errorWithin 10%
Up to €250,0002287.3%59%
€250,000 – 400.0002098.2%58%
€400,000 – 600.0001836.5%67%
From €600,0001507.0%63%

Recent sales

The test goes back to 2003, because older sales also say something about how the model calculates. On the 101 assessed sales from 2025 and 2026 the median error was 6.9%, and 70% were within 10%.

Where the model is weaker

Mainly apartments tested
690 of the 770 sales were apartments. We can say little yet about single-family houses (43 tested, 9.6% median error).
Streets with very different buildings
Where homes in one street differ greatly in size, condition or layout, the estimate is off more often. The model then usually gives low confidence itself.
Condition of the home
The Kadaster doesn't know whether a home was renovated or a doer-upper. So that was missing from the test; in a report you fill it in yourself.

What the numbers don't say

A median error of 7.2% doesn't mean your report is 7.2% off. It can be closer or further off. So always look at the confidence and the sales in the report itself.

This is an indicative valuation, not a formal appraisal. For a mortgage you need a validated appraisal report from a recognised appraiser.

See the sample report Guides for buyers

How to compare valuation tools

Many sites give a home value. Ask every tool these six questions, including us.

QuestionWoonWaarde
How big is the error, and on how many sales was that tested?Median 7.2% on 770 assessed Kadaster sales, 9.9% with the 176 differing sales included. The model was tested without human review.
Where does it work less well?Among others, at low confidence and in Amsterdam, which has been tested too little with 17 sales. It's on this page.
Can you see the sales and adjustments behind it?Yes, every sale and every adjustment in euros.
Does a person look at it?Yes, every report is checked by a person.
How does the provider make its money?Not from your bid: no commission and no referrals. The report is free during the introductory period.
How fast is it?Up to two working days, because we review every report ourselves. Many other tools are ready instantly. You do see the first indication from the WOZ value (the value your municipality sets for taxes) straight away.

The first indication with your WOZ value

On the homepage we fetch a home's WOZ value, or you enter it yourself. We add the national price change from CBS and Kadaster since the WOZ year. The range is 14% either way, and 20% for a WOZ value under €300,000. This is separate from the model above: nobody looks at it, and condition, energy label and sales in the street don't count.

81 sales, 2024–2026Result
Median error6.4%
Purchase price within the range86%
Within the range, WOZ under €300,000 (23 sales)74%
Within the range, WOZ from €300,000 (58 sales)91%
More than 20% off7%

Each sale got the WOZ value published at that time. We chose the €300,000 cut-off and the widths after seeing these sales, so these figures are probably too favourable. It is also a small test: 72 of the 81 sales were apartments, mostly in the Randstad. It says little about single-family homes and the rest of the Netherlands.

That is why the homepage says about 1 in 4 homes sell outside the range, not 1 in 7. A much larger test by Calcasa on 12,171 sales in the Groningen earthquake area found 59% within 10% and 87% within 20%. Research by the Waarderingskamer shows the WOZ value matches the market well on its valuation date. After that it falls behind prices, which is why we add the price change.

Questions about accuracy

A short answer to what we hear most often.

How accurate is an online valuation?

That varies by provider and by home. In a test on 770 Kadaster sales, WoonWaarde's model was less than 7.2% off in half of the cases, and within 10% in 62%. At high confidence the median error was 5.4%.

Why do you also publish where it goes wrong?

Because you can only use an estimate if you know how far off it can be. That's why the excluded sales, the low-confidence cases and the weak spots are also on this page.

Is the human review included in the test?

No. The figures are about the model alone. In a real report, a member of staff then checks whether the sales are really comparable and whether anything is missing.

How often are these figures updated?

We run the same test again with every change to the model. The date of the latest test is with the four numbers at the top.

Thinking of making an offer?

First, see what similar homes have sold for.

Enter the address