Automated Property Valuation: How Accurate Is It Really?

A price estimate can be delivered in seconds. That’s the easy part. The hard part is knowing how much weight you can put on it — because a number without context is as dangerous as it is useful. The question is rarely whether an automated property valuation can get it right, but when it does, and how you can tell the two situations apart for yourself.

For an agent, a valuer, a credit officer or an investor, the point is not to replace a concrete valuation with an algorithm. The point is to know when the estimate is solid enough to stand on its own, when it needs qualifying, and when you should go out and see the property with your own eyes. It comes down to reading the confidence just as carefully as the number itself.

What an automated model actually does

An automated property valuation — known internationally as an AVM (Automated Valuation Model) — estimates a sale price from patterns in past transactions. The model finds comparable properties, weights them by similarity and proximity, and adjusts for the differences between them and the property you’re looking at: floor area, age, condition, location, plot size and a range of other factors.

In principle it’s the same exercise a valuer carries out manually when comparables exist on the existing market — just performed across far more transactions and far faster. The logic behind the method is the same whether it’s done by a person or a model; we’ve described it in detail in the article on how a property is valued on the existing market.

The difference lies in scale and consistency. A model can work through thousands of transactions in an area and hold the same methodology from one address to the next. In return, it can only see what’s in the registers. It can’t walk through the front door.

Three things that determine accuracy

A valuation’s accuracy is not a fixed property of the model. It varies from address to address, and it depends on three things.

1. Data density — how many comparable sales are there?

The single most important factor is how much relevant transaction data exists around the specific property. A flat in a large apartment block in a major city typically has dozens of near-identical neighbouring sales within a short timeframe and distance. Here the model has a strong basis to work from.

An architect-designed villa on a large plot in a sparsely populated area is the opposite extreme: few sales, wide dispersion, and each property is effectively unique. Here the estimate becomes more uncertain by definition — not because the model is poor, but because reality gives it little to go on.

Rule of thumb: the more uniform a housing segment is, and the more sales there are around the property, the more you can rely on the estimate.

2. Data quality — are the registers accurate?

A model is only as good as the data it’s fed. The areas, the building’s year of construction and the use codes in BBR (the Buildings & Dwellings Register) are the foundation — and if they’re wrong, the error feeds straight through to the estimate. An incorrectly recorded living area or an unregistered extension can shift the valuation noticeably.

This is precisely why register data should never be taken at face value without a critical eye; we’ve gathered the typical traps in our review of the pitfalls in BBR data. For an automated valuation the simple rule applies: garbage in, garbage out. An estimate at an address with questionable register data deserves a second look.

3. Condition and what the model can’t see

The hardest factor to capture automatically is the building’s physical condition. Two properties with identical register data can have wildly different values if one has been newly renovated and the other is run down. Condition can move the price per square metre considerably — we’ve covered the relationship in the article on what a building’s condition level does to its value.

A model can estimate a likely condition level from age, the most recent renovation year and the character of the area, but it can’t see a new kitchen or a leaking basement. This is where the human inspection remains indispensable — and it’s one of the reasons an estimate should be read as a starting point, not a definitive answer.

The confidence level matters more than the number

This is where many people read an automated property valuation wrong. They focus on the price and overlook what sits beside it: the confidence level.

The confidence level is the model’s own assessment of how sure it is about its estimate — derived from exactly the three factors above. A high confidence level typically means many comparable sales, a uniform segment and solid register data. A low confidence level signals the opposite: thin data, wide dispersion, or an atypical property.

In practice you can think of it like this:

• High confidence: In many cases the estimate can stand on its own as a first indication — for example, for a quick screening, a portfolio review or an initial price indication.
• Medium confidence: Use the estimate as a starting point, but qualify it. Check the register data, look at the specific comparables, and pay attention to what’s pushing the uncertainty up.
• Low confidence: Treat the number as a rough bearing. Here an inspection and a manual valuation are genuinely necessary before you base a decision on it.

Reading the confidence level is what separates the professional user from someone who simply reads off a number. An estimate with low confidence isn’t a bad estimate — it’s an honest estimate, telling you that you need to do more work.

When can an estimate stand on its own — and when not?

The honest conclusion is that it depends on the purpose. For a quick indication, an initial sort of a long list of properties, or a conversation with a seller, a high-confidence estimate is often more than enough. For a credit decision, a final price, or a transaction of significant size, an estimate should never stand on its own — regardless of confidence. Here it’s an effective starting point that saves time and sets direction, but not a substitute for a concrete valuation.

It’s also worth remembering that an estimate is a snapshot of the existing market. It’s not a forecast of where prices are heading. It estimates what the property could most likely be traded for now — based on the transactions the market has actually delivered. Once you understand the price per square metre as an area-level concept, it also becomes clearer why estimates vary geographically; we’ve unpacked this in the article on how the price per square metre for an area is calculated and read.

What the exercise looks like in Arcili

All of the above — finding the comparables, weighting them by similarity, adjusting for area, age and condition, and determining how certain the conclusion is — is an exercise that takes time if done manually. It’s the exercise Arcili automates in the Boligvurdering (Valuation) module.

The valuation is built on public registers and transaction data from the existing market, weights a large number of factors per address, and states a condition level. Rather than simply giving you a single number, it returns an estimate together with a confidence level, so you can see straight away whether the basis is solid or thin — and decide for yourself whether the estimate can stand on its own or needs qualifying with an inspection. It doesn’t replace the valuer’s judgement; it gives you a well-founded starting point in seconds instead of hours.

If you’d like to see what the confidence level and the condition assessment look like on your own addresses, you can book a walkthrough and have the module shown on specific properties.