A British pub name is rarely an invention. It is assembled, from a short vocabulary of colours, animals, trades and infrastructure, and it is stubborn: the sign outlives the thing it was named after. This takes every named pub in Great Britain ( of them, from OpenStreetMap), pulls each name apart into its slots, and then measures every one against the railways and canals on the ground around it - the lines that still carry traffic, and the lines that do not. Where the trains have gone, the railway names stayed: of the railway-named pubs more than a kilometre from any working line, stand nearer the lifted trackbed than the pub next door. The same test on the Red Lions and White Harts finds nothing.
One · the grammar of the sign
Most pub names fit a template. The largest is a colour followed by an animal, which accounts for pubs across different combinations - and the grid is mostly empty, because the vocabulary is heraldic rather than descriptive. There are Red Lions everywhere and almost no Red Lambs. Click a cell to pick those pubs out on the map below.
Counts are of pubs whose name begins with the colour and animal, so The Red Lion counts and The Lion and Lamb does not. Colours and animals are hand-listed in src/names.py; anything outside those lists falls through, and the unclassified share is reported in the table at the foot of this section.
pubs end in Arms. Of those, borrow the arms of a trade rather than a family or a regiment - which is to say the sign is a record of what the people in that street did for a living.
The other productive template pairs two signs, usually because two houses merged or a landlord wanted the arms of two patrons. The pairs are not random: the second slot is nearly always the senior one.
Two · the whole board
Britain drawn in pubs. There is no coastline in this picture and no basemap of any kind - the shape is dots, one per pub, and the lines behind them are railways and canals out of the same OpenStreetMap extract, in green where trains still run and oxblood where the line has been lifted. Drag to pan, scroll to zoom, click a pub to read its sign.
Click a pub
Three · what the name remembers
pubs carry a railway name - Railway, Station, Junction, Locomotive, Signal Box, the Great Western. Britain lifted roughly a third of its network in the twenty years after 1955, so the question is whether those names now sit by lines that are gone. The difficulty is that railways and pubs are both urban: in a town, a dismantled line is near everything, and a working station is ringed by lifted sidings. Three measures are reported. The first two were the original plan, and the check meant to validate the second one failed; the third was built to get round that, and it carries a placebo of its own.
Measure one, and its weakness. Take the sites that are more than a kilometre from any working line but within 250 metres of a dismantled one, and ask what share of the pubs there are railway-named. It is a large lift, and it is also exactly the number a sceptic should distrust: those sites are disproportionately in old industrial districts, where both dismantled lines and railway pubs are common for reasons that have nothing to do with each other.
Measure two, paired with the pub next door. Every railway-named pub is paired with the nearest other pub - a median of metres away, the same street in most cases - and the two are compared on how far they sit from the nearest dismantled line. The control is in the same town, the same decade of building, the same industrial history. Whatever urbanity does to the measurement, it does to both of them.
The plan was to run the same paired test against lines that are still open, expecting no gap: if the railway pubs were no nearer working lines than their neighbours, the dead-line gap would have to be memory. The gap against working lines is larger. Of course it is: a Station Hotel stands by a station, and most stations still have trains. So measure two cannot tell memory from proximity to a working railway, whose old sidings and goods yards are themselves mapped as dismantled track. It is kept here as the check that did not do its job.
So take the working railways out of the picture. Keep only pubs more than a kilometre from any line that is still open ( of them), and pair every railway-named pub among them with its nearest neighbour that is also that far from a working line. There is no live station nearby to stand beside and no working yard shedding mapped sidings. If the railway pub still sits by the old trackbed, the name has outlived the line.
Two checks on measure three. The vocabulary is a judgement, so the test is re-run on only the two words nobody uses for anything else, Railway and Station. And as a placebo it is run on the heraldic signs, colour plus animal, which have no reason to sit by a railway: if pairing far from the working lines produced a gap by itself, the Red Lions would show one too.
Four · method
The data is OpenStreetMap, which is volunteered. A pub is here because somebody mapped it, and Britain is mapped unusually well, but coverage is not uniform and the count will not match the licensing statistics. Names are as mapped, including the occasional chain suffix and the occasional misspelling.
Distance is to the nearest line, not along a road. Lines are densified to a vertex every 40 metres and the nearest vertex is taken, so a quoted distance is within about 20 metres of the true perpendicular distance.
“Dismantled” is OSM’s word, not a historical record. It covers railway=abandoned|dismantled|razed|disused and the disused:railway=* lifecycle spellings. Mapping of old trackbeds is enthusiastic and good but not complete, which biases the finding against itself: a trackbed nobody has mapped makes a railway pub look further from a dead line than it is.
The vocabulary is a judgement. Colours, animals, trades and the infrastructure word lists are hand-written in src/names.py. They are deliberately narrow: a name matches a template or it falls through, and the fall-through share is published rather than tuned away.
Pubs mapped twice are merged. The same house often appears as both a point and a building outline; where two records within 60 metres share a canonical name, one is dropped ( merged).
Nothing is fetched while you read this, except the visit count in the bar at the top. The whole dataset is compressed into the page.
What failed. The live-line check above, which was meant to validate measure two and instead showed the same pull towards working lines, more strongly. The word Mallard, in the railway list for the A4 locomotive until the names were read and nearly every Mallard turned out to be the duck; Grand Junction, which is a canal. And the canal test, which comes out null: too few disused canals are mapped to measure against.
The paired test is a sign test, not a model. It asks only which of the two pubs in each pair is closer to a dead line, which is why it needs no assumption about how distances are distributed.