Methodology
How Realto puts a price on a home
Every estimate comes from a model trained on hundreds of thousands of Dutch residential listings. This page explains what goes in, how the number comes out, and exactly where the limits are — no statistics background required.
The band we show around every estimate.
Of estimates on listings the model had never seen.
Dutch homes in the current training set.
These figures are read from the model version currently serving estimates, and change when it is retrained.
From an address to an estimate
Four steps, and the whole thing takes about as long as loading this page.
You give an address
We locate the home and pull together what is already recorded about it — type, size, layout, build year, energy label.
You correct the profile
Records go stale. You adjust anything that is wrong before the estimate runs, so nothing is assumed that you can tell us.
The model reads it
Twenty-one characteristics go into the trained price model, which returns one number and the reasoning behind it.
You get the whole picture
The estimate, the range around it, what drove it up or down, and a ten-year outlook for the area.
What the number means
Realto estimates the asking price a home like yours would carry on today's market. Being clear about what that is — and is not — matters more than any decimal place.
A data-driven indication of what comparable homes are being asked for, under comparable conditions, right now.
A consistent yardstick — the same model, applied the same way, so two homes or two scenarios can honestly be compared.
A starting point for the conversation you are about to have with an agent, a buyer, or yourself.
A guaranteed selling price. What a home fetches depends on the buyer who turns up and the day they turn up.
An official valuation. Banks, notaries and the tax authority require a certified appraisal. Realto is not one and does not replace one.
Advice on what to list or offer. That is a strategy call, and it belongs to you and your agent.
Where the data comes from
Three sources, deliberately kept separate — one for what homes are worth, one for where the market is heading, one for where the home actually stands.
Residential listings
Houses and apartments across the country, each with its asking price and full set of characteristics. Only listings from 2024 onward are used, so the model learns the market as it is now rather than the one before it.
The official price index
The national statistics office publishes a quarterly index for existing owner-occupied homes — for the country, every province and every municipality. It anchors the long-term view.
Location context
Coordinates, municipality and neighbourhood, so the model can tell one street from another instead of working with regional averages.
All of it is used to learn structural patterns, never to look up a neighbour. No single listing decides your estimate, and there is no fixed price-per-square-metre table sitting behind any of it.
What the model actually reads
Twenty-one characteristics, in six groups. No price-per-square-metre rule is applied anywhere — every one of these is weighed against every other.
Where it is
Consistently the single largest driver of a Dutch home's price.
How big it is
Floor space matters most; land and garden matter differently in a city than outside one.
How it is laid out
Two homes of the same size are not worth the same if one splits it into more usable rooms.
What kind of building
Detached, corner, row, maisonnette, penthouse — sorted into a fixed set the model can learn from.
How comfortable
Labels are ranked G through A5 in order, so a step up is read as a step up — not as a different category.
When
Every listing carries the month it appeared, so the model can track the market moving rather than averaging it flat.
Where a characteristic is genuinely missing, it is filled in the same way every time and never invented per home. A missing energy label, for instance, is stood in for by what is typical for that build year — and you can override it before the estimate runs.
How the model works, in plain language
Realto uses gradient boosting: instead of one big formula, the model is thousands of very small decision trees, each one trained to correct the mistakes the previous ones made.
Stacked together, they pick up patterns nobody wrote down — that in one neighbourhood the jump from label D to label B is worth more than an extra bedroom, and in another it is the other way round. A fixed rule cannot hold both of those at once. A learned model can.
Categories like city, neighbourhood and building type are handled natively rather than flattened into numbers, which is what lets the model keep thousands of distinct Dutch neighbourhoods apart instead of blurring them into regional averages.
It thinks in percentages, not euros
The model is trained on the logarithm of price, so it optimises relative error. A €40.000 miss on a €1.2 million villa is not treated as the same mistake as a €40.000 miss on a €250.000 flat.
It is graded on its own future
Training stops at a cut-off date and accuracy is measured only on the months after it — listings the model has never seen. That is a harder test than a random split, and a fairer one for a market that moves.
It is retrained, not patched
New data means a full retrain and a full re-scoring. Each version's accuracy is recorded before it serves a single estimate, and the version that produced your number is stamped on it.
Why this price, and not another one
Every estimate is opened up. The model starts from a baseline — what an average home in the data is worth — and each characteristic of your home pushes that number up or down. Those pushes are calculated so that they add back up to exactly the estimate shown. Nothing is left over.
Illustrative figures from a sample estimate. Your own breakdown appears on every result page.
Why we show a range, not a single number
A single euro figure would imply a precision that no model has. So every estimate is shown with a band around it, and that band is not a guess — it is the model's measured track record on listings it had never seen.
Half of those estimates land within 5.9% of the actual asking price. That is the number the band is built from, and it is recalculated every time the model is retrained.
The rest of the distribution is published too, because a single accuracy figure hides as much as it reveals.
Testing a change before you make it
Insulate to label A. Add a bathroom. Extend by twelve square metres. Change one thing on the profile and the same model runs again — everything you did not touch stays exactly as it was, so the difference you see is that one change and nothing else.
- Living area
- 99 m²
- Bathrooms
- 1
- Energy label
- C
- Solar panels
- No
- Living area
- 99 m²
- Bathrooms
- 1
- Energy label
- A
- Solar panels
- No
A scenario shows what the market has historically paid for a characteristic — not what the work would cost you, and not a promise that a buyer will pay the difference. Treat it as the value side of the sum, and get the cost side from a builder.
The ten-year outlook is a separate model
The long view does not come from the price model. It comes from the official Dutch statistics office's index for existing owner-occupied homes, published quarterly for the country, every province and every municipality.
That index is extended forward with a damped trend — recent momentum is carried on, but it fades rather than compounding forever — and then blended toward a long-run growth rate of about 3% a year. Extrapolating a boom in a straight line for a decade is how forecasts go badly wrong.
Your home's curve is simply that path scaled by your own estimate. The shaded bands come from testing the forecast against what actually happened, so they widen the further out you look. That widening is the honest part.
What the model cannot see
Some of what makes a home worth more never reaches a dataset. These are the gaps, stated plainly, because knowing where an estimate is weak is what makes the rest of it useful.
The state of the inside
A tired kitchen and a new one are the same square metres to the model.
Character and views
An architectural one-off, a canal view, a listed façade — none of it is recorded as a field.
Micro-location
Which side of the street, how loud the road is, whether the garden gets afternoon sun.
The people bidding
Two determined buyers can move a price further than any characteristic on this page.
Sudden turns
A rate move or a policy change shows up in the data only after it has already happened.
So use it for the part it is good at
Realto narrows the question to a defensible range. Local knowledge closes the last stretch.
How current it is
Listings arrive continuously and the price model is retrained on the newer ones. The market index refreshes each quarter, as soon as it is published. Every estimate carries the model version that produced it, so a number from last month can always be traced back to the model that made it.
What we do with your address
An address and the details you enter are used to produce your estimate and your scenarios. That is the whole purpose. We do not sell personal data, and we handle it under the GDPR. The specifics are in the privacy policy.
The point is not to tell you what your home is worth.
It is to show you what the market has paid for homes like yours, which parts of yours are carrying that number, and what would move it. Then the decision is yours to make, with the reasoning in front of you.