Partner briefing · Exeter · October 2026

Where Exeter's indoor 4G gaps cost customers, and which ones are worth a mast

An independent read from open data: Ofcom drive tests, Environment Agency LiDAR and Census 2021, for 152,419 residents in Exeter and its edges.

Peter Sgithub.com/petersolan/Mobile-Signal-Strength

The answer first

Three operators, three different moves

VMO2

Build

Most exposed: 9,400 to 22,900 residents get indoor 4G from a rival but not from VMO2.

One site at the University or Central Exeter pays back in both scenarios, needing 9-18% of the customers it covers against a fair share of about 30%.

EE

Measure first

The case turns on how optimistic the model is: its best sites need 60-186% of covered customers as modelled, 20-36% bias-corrected.

A one-day drive in three areas settles it before about £750k is committed.

VodafoneThree

Don't build, market

Its gaps are small and scattered; the best site needs 41-93% of covered customers.

2,000 to 5,200 residents get indoor 4G only from VodafoneThree. That is a marketing opportunity, not a network one.

Ranges: as modelled / bias-corrected (slide 6)2

Where the gaps are

Three maps of Exeter: for EE, VMO2 and VodafoneThree, the addresses where a rival has indoor 4G and the operator does not, with recommended new sites as stars

Residents a rival covers indoors and you do not

Operator	As modelled	Bias-corrected
EE	5,935	23,189
VMO2	9,429	22,932
VodafoneThree	1,964	4,207
No operator	795	2,374

Indoor 4G is Ofcom's -95 dBm. The gaps cluster on valley sides and in the dense centre, where hills and buildings block the existing masts.

Stars: recommended sites, numbered in build order (bias-corrected scenario)3

Who the customers are

Not every gap is worth the same: the University area is the prize

Operator	Defend	Attack	Top area to defend
EE	5,120 / 19,736	864 / 1,908	University; Central
VMO2	8,419 / 19,900	72 / 335	University; Central
VodafoneThree	1,615 / 3,454	1,795 / 4,716	Exwick & Foxhayes

50% of residents around the University are full-time students (Census 2021)

Heavy data users who pick a network every September. Weighted by value (employed or studying, relative to the area), the University tops VMO2's list in both scenarios.

Defend: adults a rival covers indoors and the operator does not, the customers it loses on coverage. Attack: adults only it covers, where marketing wins. As modelled / bias-corrected.

Census 2021 via Nomis: TS066 economic activity, TS007A age, per LSOA4

The money

Share of covered customers each site must win or keep to pay back in five years

944 subscribers a £250k site must win or keep

£250k all in (site, planning, power, backhaul, build), 10% a year to run, £13.24 ARPU (Ofcom, Q4 2025), 50% margin, five-year payback.

Site cost	Subscribers
£250k	944
£500k	1,888
£750k	2,832

VMO2's Central site still pays at £750k if the bias correction holds (26% of 10,729).

Break-even instead of a guessed switching rate: no public data says how many covered customers switch5

How sure we are

The model is right about where, and optimistic about how much

Network	Indoor, measured	Indoor, modelled	Same class
EE	87.6%	98.4%	87%
O2	80.7%	89.4%	80%
Vodafone	85.5%	84.0%	81%
Three	65.5%	73.6%	78%

On drive-test areas the model never saw (one in five 1 km squares held out), it overstates indoor coverage by 8-11 points for EE, O2 and Three.

So every result comes twice

As modelled: gaps are lower bounds.

Bias-corrected: each network's signal lowered by its held-out mean error (EE 6.3 dB, O2 4.8, Three 4.8, Vodafone 3.9).

The two agree on where to build, usually within 500 m, and on the order of the operators. They differ on how much money is at stake.

Same analysis also written in SQL (DuckDB) and checked against the Python: every figure agrees6

The pilot

Spread the drive over places, not readings: the model's error is local

1 km squares driven	Right call	95% range of the measured error
1	39%	±21 dB
3	50%	±12 dB
5	58%	±10 dB
8	63%	±8 dB
12	68%	±6 dB

75% of EE's error varies between 1 km squares; readings within a square mostly agree, so 100 readings per square tell little more than 10. And the optimism is uneven, so even 12 squares get a citywide bias right only about two times in three. Worst case of the two possible truths, 30 road squares per 1 km square.

Decision rule

Drive EE's three target areas (Pinhoe, the University, Central Exeter) with a scanner and compare each 20 m road square with the model.

Do not estimate a bias and apply it. Drive every 1 km square inside each candidate site's footprint, about 10 road squares in each, and count the customers there without indoor 4G. Build where that count clears the 944-subscriber break-even at a fair share.

After a site is built

Compare covered neighbourhoods with similar ones the site does not reach, before and after: indoor coverage share, port-outs and data traffic, over three months each side (difference in differences).

Simulated from EE's held-out errors, resampling whole 1 km squares, 4,000 pilots per design7

What would sharpen it

Open data gets the direction; partner data would get the numbers

Real cell locations and antenna patterns: masts here are estimated from where the signal is strongest.
Device-side measurements: they see indoors and every street, where drive tests see only roads.
Local market share, churn and port-outs by postcode: replace the national 30% share and the break-even framing with actual switching.
Capacity: busy cells need 1,800-2,600 MHz layers on existing masts rather than new sites; drive tests measure signal, not load.

How it was built

Python and DuckDB SQL on Ofcom's drive tests (2022-2024), 1 m LiDAR and Census 2021. Path loss with terrain diffraction, calibrated per operator and validated on held-out areas.

Coverage checker: petersolan.github.io/Mobile-Signal-Strength/coverage.html

Code and method: github.com/petersolan/Mobile-Signal-Strength

Peter S8