evidence

We call it before it happens. Then we publish what happened next.

Every forecast is made before the month it describes, and measured against what the police actually recorded. Most of the 67,050 out-of-sample forecasts land above 98% accuracy.

Request the full evidence pack See the cases

Called before the fact

Every forecast is made before the month it describes. Nothing is fitted after the outcome is known.

Checked against the record

Each one is measured against what the police actually recorded, never against another Dawn output.

Across the country

Every territorial police force in England, Wales and Northern Ireland. Scotland is ready.

Accurate at rest, and accurate when harm is moving. Every case below is drawn from a month where it moved.

the hard part

The series most people cannot build

Domestic abuse is not an offence class. It is carried as a flag on recorded offences, and it does not appear in the open street-level release at all. A national, neighbourhood-grain domestic abuse series does not exist in published files. You cannot download it. You cannot reconstruct it.

That series exists because Dawn built it. Every figure here is measured against a recorded count, never against another Dawn output.

the cases

Forecasts, set against what was actually recorded

Part one Neighbourhood grain

Example area, template only

Cutoff The last month visible to the model at the point of forecast.

What the series was doing: Flat for two years, then a shallow rise across the last quarter.
What Dawn saw: A rise flagged before the month began, ahead of what the area normally sees.
What the standard view showed: The rolling twelve-month average still read the area as steady.
What was recorded: The recorded figure confirmed the rise, within the forecast band.
In people: Illustrative only. Real people-level context is supplied with the real case.
Why this happened operational prior: Example driver, stated as the reason to look and kept out of the arithmetic.
What it changes: The decision this month of warning supports, in this area, in this window.
The decision chain: Signal, to the team whose job it is to act, to the step they could take earlier.

both directions

It reads direction both ways

Dawn calls falls as well as rises. A model that reads direction both ways is modelling, not leaning on a trend.

The full set of down-calls is in the evidence pack.

Request the evidence pack

scope

What Dawn does

Forecasts where harm is building, across recorded offences, from neighbourhood to local authority grain, up to a year ahead.

Coverage

Every territorial police force in England, Wales and Northern Ireland, across the offence types that carry the most harm. Scotland infrastructure ready.

How accuracy is measured

One minus the absolute percentage error against the recorded count, measured over March 2025 to May 2026, across all 67,050 out-of-sample forecasts.

Methodology

Full methodology and denominators available on request.

See it against your own area.

The evidence pack includes the full study, the methodology and every case in full.

Request the evidence pack Book a walkthrough

Dawn was built to see harm before it lands. Read why that matters. Read the story