Methodology: Locally-Sensitive Prediction Model (Phase 5.1)
This model predicts ward-level results for the 7 May 2026 English council elections using a 6-step locally-sensitive approach that combines Census demographics, historical election baselines, party demographic profiles, and current polling.
1. Demographic profiles (data-driven, not uniform multipliers)
Using ~1,400 wards with both Census data and historical results (2024/2022), we compute each party's average vote share by:
Education quintile (% with Level 4+ qualification)
Ethnicity quintile (% non-white population)
IMD deprivation quintile (from PCON data)
Council type (London borough, metropolitan, county, etc.)
This generates a local demographic multiplier for each party in each ward: demo_mult = expected_local_share / national_average_share. This replaces the uniform national multipliers used in earlier versions.
2. Reform swing model
Since Reform barely stood in pre-2024 local elections, their demographic variation uses the parliamentary swing regression:
R² = 0.553 from 571 constituencies. This produces a multiplier applied to Reform's national poll share.
3. Polling inputs (February 2026)
Party
National poll %
Reform
29.7%
Labour
19.3%
Conservative
19.1%
Green
14.1%
Lib Dem
12.3%
Raw national polls are used directly. Local variation (e.g. Greens stronger in educated wards, Lib Dems stronger where they have historical presence) is captured entirely by the demographic profile system in Step 1 rather than uniform multipliers. This means each party's effective local share varies by ward demographics and council type.
4. Alpha blending with local history
Each ward's prediction blends the demographic model with historical results:
α = 0.5 for 2024 baseline (most recent, most relevant)
α = 0.3 for 2022 baseline (older cycle)
α = 0.0 for wards with no baseline (pure model)
Reform uses reduced α (max 0.15) since their local baseline is unreliable
5. Incumbency & contest adjustments
+3pp for the party that won the ward last time (incumbency bonus)
+1.5pp for the party controlling the council
Region/council-type adjustments: e.g. Labour +4pp in London, Reform +3pp in counties
6. Candidate coverage & normalization
CON/LAB: always stand
LD: stand if historical presence >5% or degree% >25
REF: stand if swing multiplier ≥0.3 (deterministic threshold)
GREEN: stand if degree% ≥30 or historical presence >3%
Non-standing parties' shares are redistributed proportionally. Final shares normalized to 100%.
Coverage
This model covers 1,620 wards across 106 councils. 30 councils (487 wards) whose elections were postponed to 2027 due to local government reorganisation are excluded. Coverage improvements over Phase 5 come from matching county electoral divisions (CEDs) and metropolitan slug codes to Census wards using name-matching and PCON-level demographic fallbacks.
Limitations
This is a model, not a forecast. Key limitations:
~440 wards without baseline: mainly county divisions — treated as pure model predictions with no seat-change flag
No true incumbency data: the +3pp bonus is applied to the baseline-winning party, not the actual sitting councillor
No local factors: council tax disputes, scandals, independent candidates, and party organisation cannot be captured
Boundary changes: some wards have been redrawn since the baseline election
Turnout not modelled separately: local turnout (~30%) skews older and more engaged
Reform baseline unreliable: Reform barely stood in 2022/2024 locals, so the model uses low alpha for blending
~300 wards without GeoJSON: county divisions with slug codes that couldn't be matched to lower-tier ward boundaries
Data sources
Census 2021 (ONS Nomis): TS067 (qualifications), TS021 (ethnic group) at ward level
ONS Open Geography Portal: Ward to PCON/LAD/UTLA July 2024 lookup
House of Commons Library: 2022 & 2024 local election ward results
Democracy Club: 2026 ward contest list (API)
PollCheck / ElectionPolling.co.uk: February 2026 polling averages