Gail's vs Greggs

Map of London in which each postcode unit is a dot coloured by whether Gail's or Greggs accessibility is higher. A dense red Gail's core covers central and west London, surrounded by blue Greggs territory, with isolated red outposts in affluent outer suburbs. — **The partition.** Each dot is one of 460,923 postcode units, coloured by which chain wins its bisquare-weighted accessibility score at a 1,600 m bandwidth. Gail's holds a contiguous central-and-west core plus scattered outposts — Richmond, Wimbledon, Blackheath, Chislehurst. Greggs holds everything else.

44.5%

of a 4,449-parameter neighbourhood model, recovered by two numbers derived from shop coordinates alone.

+9.9pt

still added on top of deprivation and distance-to-centre together — 27% of what those two leave unexplained.

+2.0%

price gap between a Gail's and a Greggs catchment within one neighbourhood. Real, and twenty times smaller than the raw gap.

The question, and the trap in it

The obvious question — does a Gail's nearby raise house prices? — is unanswerable with this data, and would be the wrong question even if it weren't.

Gail's is a private-equity-backed chain that sites deliberately into affluent, gentrifying, high-footfall catchments. Greggs sites onto high streets and transport nodes across the whole income distribution. Any positive coefficient is reverse causality by construction.

So the project asks something else: how good a measurement instrument is the bakery map? If Gail's picks its sites well, then the distribution of two chains ought to encode a great deal about which parts of London are desirable — not because the shops did anything, but because somebody with commercial incentives already did the analysis and voted with their capital.

That reframing changes what counts as success. A proxy is supposed to read a pre-existing signal, not create one. So the null result below isn't a failed hypothesis — it's a property the instrument needs to have.

Method

Shop locations come from OpenStreetMap (Overpass) merged with All The Places, reconciled into a slowly-changing dimension so that openings and closures accumulate as an immutable event log rather than overwriting a table. The current estate is 136 Gail's and 300 Greggs inside Greater London.

Prices are HM Land Registry Price Paid, category A, Greater London, 1995 to July 2026 — 3,724,260 transactions. Geography is the ONS Postcode Directory, August 2026, which geocodes 99.94% of those sales to a British National Grid centroid and a 2011 LSOA. Retaining the 200,755 terminated postcodes is what makes the 1990s join work: the match rate never drops below 99.89% in any year back to 1995.

The accessibility index

Accessibility uses a bisquare kernel with compact support, computed per postcode unit in EPSG:27700 so distances are straight metres:

w(d) = (1 − (d/R)²)² for d < R, else 0

Raw inverse-square was rejected: it diverges as d → 0, and a hard cap on top of it puts visible rings at radius R on the map. Bisquare decays smoothly to exactly zero, is bounded at the origin, and needs only radius queries. Two quantities come out of it:

ggi_ratio = (A_gails − A_greggs) / (A_gails + A_greggs) — which chain
total_access = A_gails + A_greggs — how much bakery at all, which is essentially high-street-ness

Where both are zero the ratio is undefined, so those postcodes carry an explicit no_bakery_nearby flag rather than an epsilon. Postcodes within one bandwidth of the harvest boundary carry an edge_truncated flag, since their catchments are cut off by the edge of the query box rather than by an absence of shops.

Registration lag turned out to be far worse than documented — the incomplete tail runs about nine months, not the two the guidance suggests. July 2026 holds 28% of a normal month's sales. Rather than a fixed window, each month is scored against the median of months 12–24 back and carries a completeness ratio; the sample keeps months at ≥ 0.90.

Result 1The causal reading is a null, and five tests say so

Four specifications, each designed to break the one before. Dependent variable is log(price); standard errors clustered at LSOA; n = 2,388,494.

Coefficient plot across four specifications, showing the Gail's-versus-Greggs coefficient falling from about 0.155 with no controls to about 0.026 once LSOA fixed effects are added. — **The collapse.** Rung 1 is the ratio alone (+0.155). Rung 2 adds property type, tenure, new-build and sale-month fixed effects. Rung 3 adds distance to Charing Cross and log(total_access). Rung 4 adds LSOA fixed effects — and **83% of the raw coefficient disappears**. Within a single small neighbourhood, +0.026 survives (se 0.0057, t = 4.6): statistically alive, economically trivial.

The effect size is +2.0%, not +5.3%

The ratio is not a smooth gradient. 45% of sales sit at exactly −1 (Greggs-only catchment) and 6% at exactly +1, with a thin middle — so a linear coefficient estimates a contrast between lumps while presenting as a slope. Entered as categories instead, with Greggs-only as the reference:

Rung 4, ratio as categories
Catchment	Coefficient	se	n
Greggs only	0 (ref)	—	1,074,473
Mixed	+0.0102	0.0044	1,174,665
Gail's only	+0.0201	0.0082	139,356

The linear form implies a Greggs-to-Gail's gap of +5.3%. The honest figure is +2.0% — the linear specification overstates it by a factor of 2.6 because it stretches a scale whose mass sits at both endpoints.

Histogram of 150 permutation coefficients centred near zero, with the observed coefficient marked well outside the entire distribution. — **The placebo.** Shuffling the Gail's/Greggs labels across the 436 real locations and re-running rung 4, 150 times. Null mean +0.0003, sd 0.0049, largest absolute draw 0.0211. The observed +0.0260 lies outside all 150 draws (p = 0.007, 5.3 null standard deviations). The effect is real — it is simply small.

The pre-dating test settles the causal reading

Take today's bakery map and use it to predict sales from a decade ago, when most of the current Gail's estate did not exist.

2015 sales: +0.0187 (se 0.0081, n = 62,154)
2025 sales: +0.0258 (se 0.0080, n = 61,410)

The difference is t = 0.63. Today's Gail's map predicts 2015 prices about as well as it predicts 2025 prices. Shops that opened after 2015 cannot have caused a price surface that already existed — which is exactly what a proxy reading a stable, pre-existing property of a place should look like.

What would have falsified the instrument, and what happened

A broken pipeline. Label-shuffle placebo — the coefficient collapsed to noise on every one of 150 draws. Survived.
A radial gradient in disguise. Added to distance-from-Charing-Cross, the bakery terms moved 38.5% → 60.1%. Survived.
Redundant against conventional variables. Added on top of deprivation and distance, 63.5% → 73.1%. Survived.
Merely finer-grained than its competitors. Aggregated to LSOA means, the increment grew rather than shrank. Survived.
Just "how many shops are nearby". Against all 47,902 OSM retail POIs, the bakery terms still added 16.9 points at matched parameter budget. Survived.

The only casualty was the causal reading — a claim the project never made.

Result 2What the ratio was hiding

A ratio forces Greggs to be Gail's mirror image, so it cannot say whether Greggs proximity is good or bad on its own. Entering the two brands separately — a presence indicator plus an intensity term each, since 45% of catchments contain no Gail's — splits two effects that were cancelling.

log(total_access)+0.0206 Retail density is worth being near. This is the high-street effect, and it is positive.

log(A_greggs)−0.0230 At equal density, a Greggs-heavy catchment is a cheaper catchment. Without the density control this reads only −0.008; holding density fixed triples it.

has_greggs−0.0396 Same direction, and significant only once density is held constant (se 0.0129).

log(A_gails)+0.0402 Intensity: how much Gail's, among catchments that have any.

has_gails+0.2141 Presence: whether there is a Gail's at all. Five times the intensity term.

Rung 3 specification, all terms jointly estimated. The same pattern holds at rung 2.

Two things fall out. First, the mechanism you'd guess — "a Greggs means a working high street, and high streets are worth being near" — is half right: the high-street effect is real but it lives in the density term, and Greggs itself is negative even before density is controlled. Second, and more interesting: Gail's presence is a marker, not a dose. One Gail's tells you almost everything; five tell you barely more. That is what you would expect of a siting decision that is itself the signal.

Result 3How much of a neighbourhood model can two bakeries replace?

The test: hold property controls and sale-month fixed effects constant, then ask what share of the gap between no location information and 4,449 LSOA fixed effects each cheap location measure recovers. Reported out of sample, on a 477,492-sale holdout fitted from 1,911,001.

Share of the fixed-effect gain recovered · out-of-sample
Specification	Params	R²	Share
Base controls only — floor	0	0.5698	0.0%
+ LSOA fixed effects — ceiling	4,449	0.8352	100.0%
+ distance to nearest station	1	0.5907	7.9%
+ IMD 2019 rank	1	0.6189	18.5%
+ generic retail access (47,902 POIs)	1	0.6518	30.9%
+ distance to Charing Cross	1	0.6720	38.5%
+ bakery ratio and total access	2	0.6878	44.5%
+ bakery terms, brands separate	4	0.7095	52.6%
+ IMD and distance	2	0.7384	63.5%
+ all four conventional	4	0.7467	66.6%
+ IMD, distance and bakery	8	0.7638	73.1%
+ all conventional and bakery	10	0.7663	74.0%

"All four conventional" = IMD rank, distance to Charing Cross, station distance, generic retail access. In-sample and out-of-sample shares agree to within 0.4 points on every row; the 4,449-parameter ceiling moves only 0.8359 → 0.8352, so it is barely overfitting at ~430 sales per LSOA.

Reading the table honestly

The headline deflates, and then stops deflating. "44.5% of a 4,449-parameter model with two numbers" sounds remarkable until you notice that distance to Charing Cross alone gets 38.5%. The number that survives scrutiny is the increment: adding the bakery terms on top of IMD and distance moves 63.5% → 73.1%, closing 27% of everything those two leave unexplained. That comparison has the obvious alternatives already in the model, so there is nothing left to deflate it against.

Two bakery chains beat the official deprivation index by roughly 3× — 44.5% against 18.5%. That asymmetry cuts both ways and the write-up says so: IMD's domains include income and employment deprivation, so it is a survey-derived measure of socioeconomic status, partly built from the same forces that set prices. It is not a neutral competitor, and it is not built to predict prices. That makes its weak showing less surprising and the bakery index's showing more so, since the latter uses nothing but coordinates.

Station distance is a genuine negative result. 7.9% alone, and +0.1 points on top of IMD and distance. London is dense enough with stations — median nearest 629 m — that the variable barely discriminates. OpenStreetMap was used rather than TfL open data, because TfL omits National Rail suburban stations, which in outer London are often the only station a postcode has; using TfL would have understated accessibility exactly where the bakery index is weakest, flattering it.

Is Gail's vs Greggs actually the best pair?

The same machinery, six retail pairs, identical sample and parameter budget, ranked by increment over the full conventional stack.

Retail pairs by marginal contribution
Pair	Outlets	Alone	Increment
Gail's / Greggs	105 / 225	51.0%	+6.7
Waitrose / Iceland	141 / 131	40.3%	+6.0
Pret / Greggs	316 / 225	36.4%	+3.1
M&S Food / Poundland	151 / 76	30.1%	+2.7
Costa / Pret	400 / 316	35.8%	+2.6
Boots / Superdrug	298 / 120	16.1%	+2.3

Harvested from OpenStreetMap only, so the Gail's/Greggs row differs from the main analysis, which additionally merges All The Places and applies a curated duplicate list. Internally consistent within this table.

It isn't outlet count

The obvious alternative explanation is resolution: a pair with 716 outlets has a finer-grained index than one with 330, regardless of how sharply the brands sort. The ranking already runs against that — Gail's/Greggs has the second-smallest estate and ranks first; Costa/Pret has the largest and ranks fifth. Thinning Costa/Pret down to the incumbent's exact counts over three seeds costs it 0.2 points of increment (+2.6 → +2.4). At identical outlet counts, the gap is roughly threefold.

Two qualifications

Waitrose/Iceland at +6.0 against +6.7 is not a commanding lead — on a different sample split I would not bet on the ordering between those two, and the gap to third place is far larger than the gap between first and second. The defensible claim is a clear top pair of two, with Gail's/Greggs at its head.

More importantly, the field was selected on the dimension it is scored on. These six were chosen because they are plausibly socially contrastive, and Gail's/Greggs was chosen first precisely because the contrast is conspicuous. Winning a field selected for the property being ranked is weaker evidence than winning a random one. The honest claim is: best of six deliberately contrastive pairs, not demonstrably optimal among all possible pairs.

Limitations

No floor area

Price Paid carries only property type, tenure and new-build. Floor area lives in EPC data, which is not ingested — so rungs 1 through 3 have weak property controls and their coefficients are inflated by an unknown amount. Rung 4 is the exception: LSOA fixed effects compare houses within the same small neighbourhood, which holds much of the missing variation roughly constant. Treat rung 4, the placebo and the bandwidth profile as the real output.

The bandwidth question is undetermined, not answered

Across varying samples the coefficient rises fourfold with radius, which looks like evidence the index becomes a central-London gradient at wide bandwidths. It isn't — that comparison is confounded by sample composition, since no_bakery_nearby runs from 82% of sales at 400 m to 0.7% at 3,200 m.

Neither version answers the question. Held to a fixed sample, the standard errors run 0.006, 0.015, 0.034, 0.068 — only 400 m and 800 m carry information, and over that range the coefficient rises from +0.019 to +0.050. Above 800 m the fixed-sample design has no power, because restricting to sales within 400 m of a bakery removes the within-LSOA variation a wide kernel needs. Undetermined, not null.

Descriptive, not causal

Nothing here identifies a causal effect of anything on anything. The claim is that a bakery map is an informative measurement of location quality, not that bakeries produce it.

Current shops, historical sales

The index uses the 2026 estate against sales back to 1995. That anachronism is real, and the pre-dating test measures its consequence directly rather than apologising for it — the index working equally well in 2015 is the evidence for reverse causality, not a defect to be corrected.

Parameter budget matters more than first claimed

Against generic retail at one parameter the bakery terms add 25.5 points; at a matched four-parameter budget they add 16.9. The conclusion holds in both cases, but the magnitude is budget-dependent and asserting invariance would have been wrong.

The picture underneath

H3 hexbin choropleth of London coloured by the Gail's-versus-Greggs accessibility ratio, with a red central core, blue periphery and black dots marking individual shops. — **The index itself.** H3 resolution-8 cells coloured by ggi_ratio at 1,600 m; black dots are individual shops. Edge-truncated and no-bakery cells excluded. The structure the regression is working with is visible here: a red core with sharp boundaries, a deep-blue periphery, and a thin transitional band — which is why the ratio behaves closer to three categories than to a continuous scale.

Binned scatter of index decile against mean log price, with a steep raw series and a flat series after LSOA fixed effects are partialled out. — **Before and after neighbourhood.** Mean log price by index bin, raw and after LSOA fixed effects are partialled out. The raw gradient spans 0.29 log points; net of neighbourhood, 4.8% of that survives, and non-monotonically. The flattening between the two series is the null in one picture.

Build notes

The pipeline is reproducible from a clean checkout and runs under a 512 MB DuckDB ceiling on an 8 GB machine — a constraint that shaped the design more than expected, and cost three OOM kills before the Price Paid backfill was staged year-by-year and made resumable.

Ingest

Content-hashed idempotent downloads with HTTP conditional requests; per-source watermarks; add/change/delete upsert against the monthly Land Registry change file; transactional runs with a write gate.

Entity resolution

Cross-source spatial matching at 50 m, stable-id matching taking precedence, and explicit split/merge rules for when one row is claimed by two observations.

Data quality

Two-sided count gates that refuse the run; alerting on id-match-rate drops rather than on exceptions; repudiable runs, since a bad run can commit.

Geospatial

EPSG:27700 throughout, BallTree radius queries, bisquare kernels at four bandwidths, H3 hexbinning, GeoPandas and contextily.

Econometrics

pyfixest for high-dimensional fixed effects, LSOA-clustered errors, Frisch–Waugh–Lovell, permutation inference, nested model comparison, held-out validation.

Stack

Python 3.12, DuckDB, uv, pytest (139 tests), pandas, pyarrow, matplotlib.

Sources

HM Land Registry Price Paid and ONS Postcode Directory (OGL v3); MHCLG Index of Multiple Deprivation 2019; OpenStreetMap via Overpass (ODbL); All The Places.