How we compare an asking price to our estimate

What this number is, and what it is not

We estimate what homes like this one are advertised for in Luxembourg, then compare that estimate with the price on the listing.

It is not a valuation of the property. Nobody from ChatHome has visited the home, seen its condition, or checked its paperwork. It is not a statement about the agency or the seller either — an asking price sits above our estimate for many honest reasons, including features nobody wrote down.

• It is a comparison, not an appraisal.
• We never call a price good, bad, fair or unfair.
• For a real valuation you need a professional who visits the home.

Where the estimate comes from

A model learns from about 14,000 Luxembourg listings, current and past. For each home it reads the facts an ad states: living area, commune, plot size, number of bedrooms and bathrooms, construction year, energy class, condition, travel time to Luxembourg City and to the nearest hub, and the amenities the ad names.

The estimate is Luxembourg-only, by design. Homes across the border are priced by different rules, so they are not used to train it and it is not offered for them.

Living area and location matter most, by a wide margin. The published estimate for a home is a range, not a single number, because a range is the honest shape of the answer.

The rule that stops it being circular

This is the part most worth checking us on. If a model learns from a listing's own advertised price, and is then asked whether that price is right, it will mostly agree with itself. We measured this: scored that way, our rental predictions landed within 0.3% of the asking price, while the model's honest error was 12.7%. It was not judging prices, it was repeating them.

So a listing is never compared against a model that learned from it. There are exactly two allowed ways to produce the estimate we compare against:

• Out-of-fold: the corpus is split into five parts by property, and each home is priced by a model trained on the other four. Two listings of the same home always land in the same part, so one can never grade the other.
• First-seen: a newly published listing is priced by the model that already existed, which cannot have learned from a home that did not yet exist.

When we say nothing at all

Silence is the normal outcome, not a failure. We publish no comparison whenever the evidence cannot carry one:

• The listing does not state a living area, or does not state what kind of property it is.
• The plot is so large that the price mostly buys land, not the building.
• The price per square metre is outside anything plausible for Luxembourg — that is a data error, not a bargain.
• The model is unsure about this particular home, usually because its ad states very little. A wide range means we do not know it well enough to judge it.
• The model's accuracy has not been measured, or is too poor to support any public statement.
• Our estimate is more than a few weeks old, because the market moves and a stored answer does not.
• The listing is a rental. Our rental model is graded on too few homes to judge somebody's advertised rent.

How big a difference has to be before we speak

Every estimate is a range that we expect to contain about half of homes like this one. Half of all listings therefore fall outside such a range naturally — so being outside it is not, on its own, evidence of anything.

We add a second and much stricter test. We measure how wrong the model is on homes it has never seen, and we take the error it exceeds only about one time in ten. A difference smaller than that says more about the model than about the price, so we call it inconclusive and show only the range.

That threshold is read from the model itself each time it is retrained. A better model starts speaking sooner; a worse one goes quiet. Nobody edits a number by hand to change what we are willing to claim.

What we still cannot do

We would rather publish our limits than let you discover them.

• We learn from advertised prices, not from what homes actually sold for. In Luxembourg, recorded sale prices run below asking prices — around 8% lower on average across the communes we can measure.
• Official per-commune sale prices exist for apartments only. There is no per-commune source for houses anywhere in Luxembourg, so house estimates cannot be anchored to recorded sales.
• The model is clearly weaker on large houses, on homes over 200 m², and wherever an ad does not state condition or energy class.
• Condition is the biggest single gap. Most listings never state it, and a renovated home and one needing work can be advertised identically.