The postcode that did not exist yet
1 in 40 sales carrying a postcode that was not in service on the day the sale completed
Every transaction in our dataset matched a postcode. That is a claim about string equality, not about time. Once you ask whether each postcode was actually in service on the day of its sale, 755,475 of 29,604,663 sales — one in forty — turn out to be carrying geography from a different period of that postcode's life. In 1995 it is one in ten. This technical note measures the problem, explains why it is administration rather than error, and states what it does and does not undermine.
- A complete join is not a correct one. Every retained transaction carries a canonical postcode that matched the directory — nothing is missing. But for 2.55% of full-market sales, the postcode's own introduction or termination date says it was not in service when the sale completed.
- The error is almost entirely historical: 9.9% of 1995 sales against 0.23% of 2025 sales. The curve falls by a factor of forty across the window and flattens out around 2009, which bounds small-area geographic findings in the 1990s far more tightly than in the 2010s.
- It is administration, not error. A unit postcode's introduction date records when it entered the current directory, so Royal Mail reorganisations stamp long-standing addresses with new, later dates. The houses did not move. Many of the addresses did not change. The postcode did.
- It kills a proposed use. The enrichment plan listed new postcodes as a proxy for new development; over 1995–2023 the correlation with new-build sales is r = 0.50 and the shapes diverge exactly where it matters. A series contaminated by reorganisation cannot be read as a construction series.
- It vindicates an earlier decision. Keeping terminated postcodes rather than using a live-only directory is what lets 39,080 sales match at all, and their median price is £50,000 against £157,500 lake-wide. Dropping them would have been a silent deletion correlated with price.
Extended findings
Three states, not two. Comparing each sale’s date against its postcode’s introduction and termination months splits the lake three ways: 28,849,188 sales on a postcode that was live (97.45%), 751,987 sold before the postcode was introduced (2.54%), and 3,488 sold after it was terminated (0.012%). The asymmetry is the tell. Selling after termination is genuinely rare; selling before introduction is not, because “introduction” does not mean what a reader assumes it means.
Why the directory looks like it is running backwards. Between 1995 and 2000 the directory records 325,818 introductions. Between 2020 and 2025 it records 97,450 — under a third, in an era of considerably more housebuilding. 2006 alone terminated 57,920 postcodes against 32,080 introduced. That is not demolition. It is administration, and it is the same curve as the anachronism series seen from the other side.
What this does and does not undermine. It does not undermine the postcode join: the match is exact, and what is at stake is the vintage of the geography attached, not whether geography was attached. It does bound small-area findings in the 1990s, where one sale in ten may carry a boundary the property did not sit inside at the time. And it removes new-postcode counts from the list of usable development proxies, which is a plan changing in response to a measurement rather than a measurement being made to fit a plan.
The limits of the measurement itself. An anachronistic postcode is not necessarily an error — this is a bound on a source of error, not a list of mistakes. The directory begins in January 1980, so earlier postcodes carry that date and pre-1995 introductions are left-censoring rather than events. Comparison is at month precision, so same-month sales count as live, which makes every figure here a lower bound. And only one directory vintage is retained: the note never claims to know what a 1997 directory would have said.
The full note carries the method — including why this is one of exactly two places in the pipeline where an as-of join is permitted — the charts behind each series, and the five caveats in full.
Use the button below to request the full document — it arrives through your in-app messages from the site administrator.