System 18 · Both halves of the join are open and keyless

Baseline

Sewage spills, adjusted for the weather

Spills fell 35% in 2025. The Environment Agency says the year was dry. Nobody has separated the £22bn of investment from the rainfall — and a new regulator is being built right now that will need to.

Climate & energyPublic moneyOperations

The problem

Every storm overflow in England now has a monitor. The 2025 return covers 14,302 overflows, and 100% of them carry a grid reference — most to one-metre precision. That is an unusually complete national dataset, free and open.

It reported 291,492 spills in 2025 against 450,398 in 2024 — a 35% fall, with duration down 48%. That is the headline the sector has been running on.

The Environment Agency’s own commentary concedes the problem with it. Its account of the year notes rainfall was "below average, with August recording just 62% of the UK's typical rainfall" and calls 2025 "a drier than average year". It attributes the improvement to weather — and then publishes no weather-adjusted figure.

Nor does anyone else. The economic regulator normalises environmental performance by sewer length, not rainfall. The charities and campaign groups that map spills do not adjust them either. So £22.1bn of committed investment is being justified against a number nobody has separated from the weather.

And the raw material to do it is sitting there. The Agency runs 1,044 rainfall stations at 15-minute resolution, free and without a key, alongside the overflow data it publishes itself.

100%
Monitored overflows carrying a grid reference — most to one-metre precision
35%
Reported fall in spills in a year the Agency itself calls drier than average
£22.1bn
Committed investment justified against an unadjusted number
0
Published rainfall-normalised spill metrics, by anyone

The system

Baseline computes what the sector cannot currently state: how much of the change in spill performance is investment, and how much is rain.

It joins each overflow to its local rainfall record and models expected spill behaviour against actual rainfall, producing a normalised performance figure per overflow, per site, per company and per catchment. The headline stops being "spills fell 35%" and becomes "spills fell 35%, of which this much is weather and this much is not."

It also fixes an identity problem the Agency created and has not solved. The published data carries a field named old_unique_id_pre_2024, because the overflow identifier scheme changed in 2024. That single break is why almost nobody publishes per-overflow trends — multi-year analysis requires resolving the same physical asset across two identifier systems, which is exactly what the entity spine does.

And it captures what is being thrown away. The statutory near-real-time feeds publish only the latest event — a start time, an end time, and current status. There is no history and no archive. The law compels water companies to publish, not to retain. Anyone recording those feeds continuously builds an event-level spill history that cannot be reconstructed afterwards by anyone, at any price.


Worked examples

Two situations this system answers

EXAMPLE 1
The problem

Spills fell 35% in a year the Environment Agency itself describes as drier than average. A regulator being stood up now must decide whether £22.1bn of investment is working, and has no weather-adjusted figure to judge it by.

What Baseline does

Baseline compares spills against how much rain actually fell nearby, using 1,044 free rain gauges — separating what the money achieved from what the weather did.

EXAMPLE 2
The problem

An overflow’s recorded spill count halves — because its monitor was offline for four months. The improvement is an artefact, and the published figure does not say so.

What Baseline does

Baseline shows how long each monitor was actually working next to every spill count, so an unwatched overflow is never mistaken for a quiet one.


Data foundation

Every dataset below is open, or its access constraint is stated

DatasetPublisherWhat it provides
Storm overflow annual returnEnvironment Agency14,302 overflows, 48 fields, grid reference on 100%. Free bulk download and open feature service, no key. Verified 72,168 records across years.
Rainfall monitoring stationsEnvironment Agency1,044 stations at 15-minute resolution. Open interface, no key, no registration.
Rainfall daily archiveEnvironment AgencyOne national file per day. Roughly fourteen months of retention — older data must be captured as it passes.
Near-real-time overflow feedsWater companiesStatutory since January 2025, published per company as open services. Current state and latest event only, with no history retained.
Water body classificationEnvironment AgencyCatchment geometry and ecological status, free and bulk downloadable — what the spills discharge into.
Companies HouseCompanies HouseResolving operators and their group structures, for the ownership side of performance.

Capabilities

01
Rainfall-normalised performance
Spill behaviour adjusted for actual local rainfall, per overflow and per company — the metric the sector lacks.
02
Weather versus investment decomposition
How much of a reported improvement is attributable to spend, and how much to a dry year.
03
Cross-year identity resolution
The same physical overflow tracked across the 2024 identifier break, which is why per-overflow trends are largely unpublished.
04
Event history capture
Continuous recording of the near-real-time feeds, building a spill history that the statutory duty does not require anyone to retain.
05
Monitor reliability
Overflows whose monitors were offline for material parts of the year — a spill count from a monitor that was not watching is not a low spill count.
06
Catchment and bathing water impact
Spills joined to the waters they discharge into, and to designated bathing and shellfish waters.
07
Regulatory reporting pack
Performance evidence in the form a new regulator would need to judge companies on day one.

Benefits

For government

  • Gives the incoming water regulator a defensible performance metric at the moment it is being stood up, when it has none.
  • Lets investment be judged on effect rather than on a figure the Agency itself attributes to weather.
  • Identifies overflows whose apparent improvement is a monitor outage rather than a reduction in spilling.
  • Provides an evidence base for enforcement that survives the obvious challenge — that a wet year would have looked different.

For the public

  • A spill figure that means something, rather than one that moves with the rain.
  • Communities can see whether the sewer serving them has genuinely improved.
  • Bill payers funding a very large investment programme can see what it bought.

Delivery

Likely sponsor
The incoming water regulator, Environment Agency, Defra, water companies
Procurement route
SBRI for the normalisation method, then G-Cloud for the service

Phasing

1
Capture and archive
1 month
Begin recording the near-real-time feeds immediately — every day not recorded is permanently lost.
2
Identity resolution
3 months
Overflows resolved across the 2024 identifier break, producing the first per-overflow multi-year series.
3
Normalisation model
4 months
Rainfall-adjusted performance, with method published openly and back-tested against wet and dry years.
4
Regulator engagement
6 months
Delivered to the new regulator and the Agency as a candidate performance metric.

Risks & mitigations

Attribution is genuinely hard

Rainfall is not the only confounder — groundwater, catchment characteristics, sewer condition and population all matter. The model must publish its assumptions and its uncertainty, and be honest that it narrows the question rather than settling it.

The regulator may build it

A new regulator with a statutory duty may develop its own metric. That would be the right national outcome, so the work should aim to inform the method rather than to be the only holder of it.

Archive decay cuts both ways

The rainfall archive retains roughly fourteen months and the live feeds retain nothing. That is the moat, but it also means the historical baseline can never be improved retrospectively.


Sources

All sources checked in August 2026. Figures carry the reference period of their source, which may differ from publication date. Where a figure could not be verified against a primary source it is not used.