How many sales does a local price figure need?

Every local property figure you read is a sample statistic, and it stops meaning what you think it means somewhere. A median built on six hundred sales describes a market. A median built on six describes six houses. The interesting question is where the line falls and what an honest publisher does on the wrong side of it.

Why does a local price need a minimum sample at all?

Precision. A median from a small sample bounces. The interval around it widens roughly with the square root of the sample, so halving the sales widens the interval by about forty per cent, and at small n the figure can move several per cent between refreshes on no real change at all.

Representativeness. Small samples are not merely imprecise, they are selected. In a quiet area, the handful of homes that did sell may be the ones that had to, which is a different population from the housing stock.

Disclosure. As the sample shrinks, a published statistic converges on a statement about individual transactions and the people in them. This is why national statistical offices suppress small cells, and in at least one country it is a condition of the licence rather than a matter of taste.

Where do the three markets draw the line?

Floors are conventions rather than theorems, and the three registers on this network arrived at comparable ones by different routes: a median needs a few dozen sales before it is published at all, and a percentile band needs more than a median does, because a quantile in the tail of a distribution is estimated from many fewer observations than the one in the middle.

Three registers, three legal regimes, three teams, and all three landed on thirty for a median. That is not coordination; it is roughly where the interval around a median stops being wider than the differences anyone wants to read from it.

The second column is where they part company, and it is also where the arithmetic is least intuitive. England & Wales publishes a median for 2,233 of the 2,281 districts where one can be computed at all, but a percentile band for only 2,004 of them: 233 districts have enough sales for a middle and not enough for an edge.

That second point is the one publishers most often get wrong. It is common to see a median published beside a “range” on the same sample, when the range needs substantially more data to mean anything. On this network a district can publish a median with no percentile band beside it, and that is the honest outcome rather than a missing feature.

Why is the French floor also a legal question?

France’s reuse conditions require that published data must not permit indirect re-identification of the people involved. A median computed from a handful of sales in a small commune moves towards being a statement about one transaction, so the floor there is doing legal work as well as statistical work.

That has a consequence people find counterintuitive: expanding coverage by widening the time window is not a free improvement. A commune with thirty sales across five years is thirty sales, however the window is drawn, and publishing it moves the risk in the wrong direction in exactly the population where it is highest. How France records a sale sets out the condition in full.

What should a publisher do below its own floor?

Four options, in descending order of honesty. Say nothing and explain why. Widen the window and say that you have. Fall back to the parent area and name it. Or publish anyway, unmarked, which is what most sites do and is the reason you can find a confident-looking price per square metre for a hamlet where four houses sold.

The test for a reader is simple: find the sample size. A publisher that shows a local figure and will not tell you how many sales are behind it has made a choice, and it is not a choice in your favour. Every figure on this network prints its own denominator, because a count beside a figure is that figure’s own quality statement. The three floors described above are visible in the products themselves: IndexProp FR leaves thousands of communes blank rather than publish a thin median, and RealScout withholds roughly a third of the medians it could compute. Registration lag is the other constraint on how local a current figure can be.