VEHICLE THEFT & SECURITY · OPEN DATA LAB
Where does vehicle theft persist?
Compare recorded offences with the vehicle population. Explore regional differences, inspect the London pilot and test whether earlier experience carries into the next year.
England & Wales · 10 regionsLondon · 32 boroughsObservation period · 2019–2025 regions; 2021–2025 boroughs
Published: crime 22 October 2026; stock 23 September 2026 · Retrieved: 26 October 2026 · Lab version: vehicle-theft-v1
Research scope: all motor vehicles, separate offence categories and licensed-stock exposure proxies. These are not insured claim frequencies or make/model security ratings.
METHOD · SOURCE COUNTS TO EXPOSURE-BASED COMPARISONS
Understand the denominator before using the result
01 · COUNTKeep offences separate
Theft or unauthorised taking (48), theft from a vehicle (45), aggravated taking (37.2) and interference (126) retain their own series. Counts are recorded offences involving motor vehicles, not just passenger cars, unique vehicles or paid claims.
02 · EXPOSUREUse vehicle-years
Each quarter contributes the average of its opening and closing licensed-vehicle stocks, multiplied by one quarter of a year. Four quarters form annual vehicle-years. Stocks include all body types, private and company keepers.
03 · COMPARETest another year
Predict 2024 and 2025 from earlier observed rates, comparing previous-year and pooled three-year experience against a common-rate baseline. Retain every diagnostic, including results that do not favour the more complicated method.
Annual rate = recorded offences ÷ estimated licensed vehicle-years × 1,000
Exposure assumptions, uncertainty and appropriate use
- Numerator geography is where the offence was recorded. Denominator geography is the registered keeper’s address. Visitor traffic and company registrations can materially affect comparisons.
- The numerator can involve SORN, unlicensed and visiting vehicles. The denominator is licensed stock. The measure is therefore a geographic exposure proxy, not the probability that a resident’s insured car is stolen.
- DfT stock values are published in rounded thousands. We retain that rounding. Vehicle-year totals derived from those values are approximate, even when displayed as whole numbers.
- Borough stock excludes vehicles under disposal or with unknown keepership. Geographically unallocated stocks are not distributed across regions.
- The 95% count-based intervals use the exact Poisson method with exposure treated as fixed. They omit reporting error, stock-estimation uncertainty, spatial dependence and systematic bias.
- Missing counts or quarter-end stocks prevent a rate from being calculated. Zero is only used when the source explicitly supplies a zero; absent rows are never filled as zero.
- National coverage here means England and Wales regions, excluding British Transport Police. Scotland and Northern Ireland are available separately as count-only modules; they are outside these exposure-based rates.
- Any pricing use requires insurer claims calibration, adjustment for vehicle/portfolio mix and testing for incremental value over existing geographic factors.
Recording changes and temporal validation
The Home Office notes revisions to Metropolitan Police 2023/24 vehicle-crime data after the move from CRIS to CONNECT and closure of dormant records. This lab consistently uses the current revised snapshot. Holdouts are retrospective comparisons using known test-year exposure, not an as-at historical backtest.
The London pilot preserves unassigned offences separately; it does not allocate them to boroughs. Their changing share means apparent borough improvement can partly reflect changes in recording. Regional London includes the full Met and City of London counts, including unassigned Met records.
The explorer uses complete calendar years: 2019–2025 for regions and 2021–2025 for boroughs. Quarterly downloads also retain the available 2026 observations. No partial year is compared with a full year.
Reproduction and release checks
Raw sources are pinned by URL, byte count and SHA-256 hash. Retained normalized inputs, parser scripts and the build script are versioned in the source repository. The build validates unique keys, offence codes, force membership, complete quarters, exposure joins and geographic coverage before replacing outputs.
200 London reconciliations passed. Borough, City and unassigned counts reconcile exactly to their force-level source values for every available quarter and offence category. These are internal consistency checks, not independent estimates of actual crime.
Rebuild command: python scripts/vehicle-theft/build.py --output-dir /tmp/theft-verify. This is a frozen release; no unattended refresh is connected. New source vintages require a fresh audit and a versioned release.
Explore vehicle populations →Compare geographic road-risk evidence →Read the parking research →