Our estimate is a range, produced by a model that learns from Luxembourg property adverts and public data. This page explains how it is built, how well it actually works, and where it falls short.
Published: August 16, 2026
We estimate what a home like yours would be advertised for in Luxembourg today, and we answer with a range rather than one number.
It is not a valuation. Nobody from ChatHome has been inside the home, seen its condition or read its paperwork. A surveyor, a bank valuer and a notary each do a different job, and this replaces none of them.
A model learns from Luxembourg homes, for sale and to rent, current and past. Each one contributes the facts its advert stated, plus the public data we can attach to its address.
Past adverts count as much as current ones. A home that has left the market still tells us what that kind of home was asked for, and when.
Homes outside Luxembourg are left out on purpose. Prices across the border follow different rules, so one model trained on both would be worse at each.
Every home is turned into a fixed list of facts. Anything an advert does not state is left empty rather than guessed — an unstated energy class is not a bad energy class.
Three models are trained rather than one. They predict a low, a middle and a high figure for the same home.
The band between them is then corrected against homes the model had never seen, so that it holds the share of homes stated in the card above. That share is deliberately not 100%: a range wide enough to always be right would be too wide to be useful.
A wide range is a real answer, not a failure. It means the facts stated about this home leave genuine room for doubt, and we would rather show that than hide it behind a confident single number.
The tempting mistake is to score a model on the homes it learned from. It then mostly agrees with itself. We measured exactly that on rentals: scored that way the model looked 0.3% off, while its honest error was 12.7%.
So every figure on this page is measured on homes the model did not train on. The data is split into five parts by property, and each home is scored by a model trained on the other four. Two adverts for the same home always land in the same part, so one can never grade the other.
We would rather publish our limits than let you find them.
The model is retrained regularly, and every figure on this page comes from the one answering right now. Its version and training date are in the card above.
When it improves, we say more on listings; when it gets worse, we say less. No accuracy figure here is ever edited by hand.
All figures below are measured on homes the model never trained on, and they come from the model serving right now.
| Kind of home | Typical miss | 9 times in 10, no worse than | Homes measured |
|---|---|---|---|
| Apartments | 9.1% | 31.1% | 7,078 |
| Houses | 12.4% | 42.5% | 4,871 |
| Studios | 9% | 37.2% | 302 |
| Kind of home | Typical miss | 9 times in 10, no worse than | Homes measured |
|---|---|---|---|
| Apartments | 12.8% | 38.3% | 935 |
The third column is the error we go beyond only about one time in ten. It is also the bar an advertised price has to clear before we will say anything about it on the listing itself — which is why a house has to differ from our estimate by far more than an apartment before we comment.
This method is published so it can be argued with. If you work in Luxembourg property and think we are missing something the market obviously prices, or that a rule here is wrong, we want to hear it. That is how it improves.
Email us about the method