The Artificial Wasteland · 31,525,946 sales · England and Wales, 1995–2026

The notch that moved

House prices pile up on round numbers, and £250,000 is about as round as a number gets. It was also, until December 2014, the point at which a British buyer’s stamp duty bill jumped by five thousand pounds in a single step. Two explanations, one spike, and no way to tell them apart, unless the tax moves and the number stays where it is. It did, nine times.

Here is the whole argument in one sentence, and the rest of this page is the apparatus for checking it. A spike at a price tells you people like the number. A void immediately above it tells you they were running from something.

1. The instrument

Every registered sale of residential property in England and Wales since 1995, as HM Land Registry publishes it: 31,525,946 transactions in a 5.5 GB file. Below is the count of sales at each £500 of price, in a £30,000 window, for one year at a time. Move the year. Change the price you are standing on.

Sale prices around the selected value A histogram of transaction counts in £500 bins for £15,000 either side of the selected price, for the selected year. The bar at the selected price is highlighted; bars above it are shaded when the tax schedule of that year put a notch there.
a notch that year (the whole bill jumps) a kink that year (only the slope changes) no threshold that year

2. What the hole is, and why it is the honest half

Bunching at a threshold is a well-worn measurement and it has a well-worn problem: the thresholds legislators pick are round numbers, and round numbers attract prices all by themselves. The literature handles this by fitting a smooth counterfactual to the distribution and calling the surplus at the threshold excess mass, with statistical machinery to keep the fit honest. That works, but the answer depends on the fit, and the confound sits directly under it.

The hole does not need a fit. Under the slab system, a purchase at £250,001 paid tax on the whole price at 3% rather than 1%: a bill of £7,500 instead of £2,500, for one pound more house. Nobody who could avoid that region stayed in it, and the count of sales in the £3,000 above the threshold, divided by the count in the £3,000 below, is a raw ratio of two integers. No polynomial, no bandwidth, no bin width. Here is that ratio for every year of the record, at the values that have been thresholds and at round numbers that never have.

Sales in the £3,000 above each price, over sales in the £3,000 below. Darker is emptier. A cell is outlined where the tax schedule of that year put a notch at that price, and dotted where it put a kink. England only, standard price-paid records. Hover or focus a cell for its counts.
Point at a cell above.

3. Four dates, read from 180 days either side

The schedule changed on days published in advance. On each of them a threshold arrived somewhere and left somewhere else, and the two round numbers used as controls were not touched at all.

Odds ratio is the change in (above ÷ below) across the date. A ratio below 1 means the region above the price emptied out; above 1 means it filled in. The z is on the log-odds scale with the usual four-reciprocal standard error, computed from the four counts shown. The difference-in-differences column divides by the same ratio at £200,000 over the same two windows.

4. The five-thousand-pound test

On 23 March 2006 the nil-rate threshold moved from £120,000 to £125,000. Both are round. Both are multiples of £5,000, £10,000 and £20,000. One of them was the threshold before that Thursday and the other was the threshold after it, and nothing else about either number changed. If round numbers make the spike, nothing should happen. If the tax makes it, the two numbers should swap.

Share of nearby sales at exactly £120,000 and exactly £125,000, before and after 23 March 2006

5. Kinks do not bunch, which is two broken predictions and one lesson

On 4 December 2014 the slab became a slice: each rate now applies only to the portion of the price inside its own band, so the bill is continuous and only its slope changes at a threshold. Two later episodes put a nil-rate band somewhere new under that system, and neither of them produced anything.

6. The bill itself

Type a price and pick a date. The schedule is parsed from the dated rate tables HM Revenue & Customs publishes, , not typed in by hand, and the staircase you see below the 2014 reform is the reason for everything above.

Stamp duty bill against purchase price on the selected date

7. What was written down first, and what happened

Nine predictions were committed to this repository before the analysis script existed, in PREREGISTRATION.md, while the pass over the 5.5 GB file was still running and nothing had been read from it but row counts. Four of them broke. The four that broke are the more interesting half.

8. Somebody got here first

The headline comparison on this page, bunching at £250,000 before and after the 2014 reform, measured on Price Paid Data, has been done. David Mesa-Ruiz (University of Edinburgh) reports it in The regional heterogeneous effects of property transaction taxes in England and Wales, a working paper whose abstract says the slab schedule “produced markedly different responses in London and Outside London”, and whose Table 3 gives excess mass at £250,000 of 2.26 in London and 1.64 outside it under the slab, against 0.39 and 0.16 under the slice. In his words, the bunching “falls sharply but does not disappear”. Daniel Borbely made the same comparison for Scotland in Fiscal Studies in 2021 and found LBTT’s kinks produce bunching that “is small and only occurs at a subset of thresholds”. Neither of those was known here when the predictions were written, and both were found by looking for them on purpose before publishing.

What this page adds is a different instrument and a longer record. The hole ratio needs no counterfactual to estimate and no bin width to choose, which means it cannot be argued with on those grounds; and read across calendar years it catches the switch happening four times, at four different prices, in both directions, with eight round-number controls that never do it. It also carries two null results on kinks that appear not to have been measured anywhere: the £500,000 nil-rate band of the 2020–21 holiday, and Wales’s £180,000 Land Transaction Tax band from 2018.

The older measurement, reproduced on a file anyone can download

Best and Kleven’s 2018 paper in the Review of Economic Studies measured these notches on “administrative data covering the universe of stamp duty returns in the UK”, obtained through the HMRC Datalab, which no member of the public can open. Run on their terms (£5,000 bins with the threshold on the correct side of the edge, a fifth-order polynomial, round-number fixed effects at multiples of £10,000, £25,000 and £50,000, and the excluded range above the notch widened until the missing mass matches the excess), the same quantities come out of the public file.

Which side of the edge the threshold sits on, and why it is not a detail

A bin is a half-open interval and something has to decide which end is closed. Put the threshold at the bottom of its bin and the pile of sales at £250,000 is counted together with the empty stretch above it, and the two cancel. Same data, same estimator, same £5,000 width, one choice of edge:

The seven notches, on the registered estimator

For completeness, the estimator this study registered before looking (£500 bins, degree four, a symmetric excluded range) against the seven figures Best and Kleven publish. The levels are not on one scale, for the reason just shown; the ordering is.

9. Check it yourself

Nothing here needs this repository. Four commands, an empty directory and about ten minutes of download.




The heavy file is 5.5 GB and the census pass takes roughly twenty
minutes on a modest machine. Everything after it runs in seconds, from the small tables
the census writes.



The five programs, at their own paths in the workspace, each served at
/checks/ plus that path:





10. What this cannot tell you





Sources