How We Predict the 2026 Local Elections
Last updated 12 March 2026
What are you predicting?
Seat counts and council control for all 136 English councils voting on 7 May 2026: 32 London boroughs (all-up), 32 metropolitan boroughs, 17 unitary authorities, 50 district councils, and six county councils. Around 5,000 seats are being contested, but because councils that elect by thirds retain sitting councillors from previous years, the model tracks all 7,300 seats across these councils to predict overall control.
Predictions are made ward by ward - a winner in each one, aggregated to council-level totals - with 80% confidence intervals and a probability of control changing hands.
How does it work?
The starting point is the last result in each ward. If Labour won 45% of the vote in a ward in 2022, that 45% is the baseline. Current polling tells us Labour is down about 15 points nationally since then, so the model marks down that ward's Labour share accordingly. The same calculation runs for all six parties in all 7,000+ wards. Whichever party finishes top wins the seat.
Reform had almost no local candidates before 2025, so most wards have no Reform baseline to swing from. Westminster polls systematically overstate the major parties and understate smaller ones in locals. And ward results are volatile enough that any single prediction needs an uncertainty range.
Why ward by ward?
National vote shares are a poor predictor of local seats. A party on 25% nationally might win 400 seats or 800 depending on where those votes are concentrated, which parties bother standing, and whether the local councillor has any name recognition.
The standard alternative is the "cube law" - take each party's national vote share, apply a mathematical formula to convert votes to seats at the council level. Most forecasters use some version of it. The cube law treats every ward in a council as identical. It can't distinguish a safe seat from a marginal.
The prediction pipeline
Convert Westminster polls to local equivalents. Look up every ward's last result. Apply swing with adjustments for demographics and incumbency, calibrated from 167,000 ward results. Filter out parties that aren't standing. Pick winners. Run it a thousand times with random noise for confidence intervals.
Data sources
- Polls: Every published poll from BPC-registered UK pollsters, compiled at pollcheck.co.uk
- Past ward results: English local election results at ward level from 2006-2024, compiled by Democracy Club and the Elections Centre (the DCLEAPIL dataset)
- 2024 general election: Ward-level estimates from Britain Elects, mainly used for Reform's geographic spread
- By-elections: Council by-election results from 2024-2025, providing the freshest available local voting data
- MRP constituency model: PollCheck's demographic swingometer covering 632 GB constituencies, used for estimating Reform and Green vote shares in wards where they haven't stood before
- Census 2021: Age, education, ethnicity, housing tenure, and EU referendum leave vote at constituency and council level
- Candidate lists: From Democracy Club, when available (usually from April)
- BBC Projected National Share: PNS estimates from Curtice and Fisher for 2022-2025, used to calibrate Westminster-to-local conversion
Reform UK
Reform contested almost no wards before 2025, so there's very little local history to work from. We use a four-tier system, choosing the best available data for each ward:
Tier 1 By-election results. Where a recent council by-election has been held in the ward, we use the actual Reform vote share - the most reliable data we have.
Tier 2 General election data. The 2024 general election results mapped to council ward boundaries. This gives Reform's geographic distribution - strong in Leave-voting working-class areas, weak in university towns and diverse cities. We convert from Westminster to local equivalents.
Tier 3 MRP constituency estimates. Our demographic swingometer produces per-constituency Reform estimates. We use the variation across a council's constituencies to differentiate within it - a ward in a high-Reform constituency gets a higher target than one in a low-Reform constituency.
Tier 4 National entry model. Reform's likely vote share estimated from area demographics. In high-Leave areas, Reform draws more from Labour; in low-Leave areas, more from the Conservatives.
County councils last elected in 2021, when Reform had no local presence. For these, we calibrate the MRP estimates against what actually happened in comparable southern county elections in 2025.
Greens and Lib Dems
The Greens and Liberal Democrats don't stand everywhere. Many wards showing them at 1-2% in baseline data won't have a candidate on the ballot.
Before April (when actual candidate lists are published), we use a proxy: if a party's baseline in a ward is below 3%, we assume they are not standing and zero out any positive swing. This is imprecise, particularly for the Lib Dems who target specific wards. But assuming every party stands everywhere produces worse results.
For the Greens, where our MRP model shows strong council-level support but a ward has no Green baseline, we inject a Green entry estimate drawn mainly from Labour, Lib Dems, and Others.
Independents and residents' groups
A popular local figure deciding to stand can swing a ward by 20 points, and no national model can see that coming.
We handle them as "Others" with a council-level floor. In councils with a strong tradition of independent candidates, wards with very low Others baselines get a floor reflecting the council-wide pattern. Without it, the model zeros out independents in wards where they regularly contest.
Where a specific local party dominates a council - the People's Independent Party in Castle Point, Aspire in Tower Hamlets, the Loughton Residents Association in Epping Forest, about a dozen others - we identify them by name rather than lumping them into "Others." Any named party with three or more councillors in the baseline gets its own label.
Independents are undercounted by 200+ seats in every backtest year.
Technical detail
Election cycles. English councils use different schedules: all-up (all seats at once), by thirds, or by halves. In 2026, most councils have all seats up - either by schedule (London boroughs, counties) or because boundary changes triggered an all-out election (Swindon and Milton Keynes normally elect by thirds but have new ward boundaries). The model predicts all wards and reconciles against the known seat count.
Boundary changes. Where a council has redrawn its wards since the last election, old results don't map cleanly to new boundaries. If the DCLEAPIL dataset has results under the new boundaries (from a by-election or partial election), we use those. Otherwise we fall back to GE2024 ward-level data (which uses current boundaries) with Westminster-to-local conversion. ONS geographic identifiers bridge ward boundaries to constituency boundaries for demographic calculations.
Demographic sensitivity. Swing is not uniform. A Labour drop hits harder in university towns than in diverse inner-city wards. We use demographic coefficients from our MRP constituency model - Census 2021 age, education, ethnicity, housing tenure, EU referendum vote - to produce a vote sensitivity modifier for each party in each ward. The modifiers operate at constituency level, so a council like Birmingham captures the variation between Hodge Hill and Edgbaston.
Westminster-to-local conversion. Local and Westminster voting behaviour differ systematically. Labour and Conservatives tend to underperform their Westminster polling in locals - people are more willing to vote against major parties when it's not a general election. Lib Dems, Greens, and Reform all overperform. We calibrate conversion factors from the BBC's Projected National Share (the Curtice/Fisher estimate each May), with additional adjustments by council type. Counties see higher Reform and lower Labour than London boroughs.
Monte Carlo simulations. We run the model 1,000 times, each time randomly varying polling error (calibrated to historical miss, with correlated shifts between parties), Westminster-to-local conversion factors, and ward-level noise (capturing local factors no national model can predict). The results produce 80% confidence intervals and a probability of control changing.
Council control. A party holding more than half the seats has majority control; otherwise it's No Overall Control. The probability of control change is the fraction of 1,000 simulations where the controller differs from the current one.
Multi-member wards. Some wards - particularly in London and metropolitan boroughs - elect two or three councillors. The dominant pattern is slate voting: people pick multiple candidates from the same party. Where the leading party has a strong vote share and a comfortable margin over the runner-up, the model gives them all seats. In tight races where two parties are close, seats are split between them - one each in two-seat wards, or two-and-one in three-seat wards. Below certain thresholds, seats are allocated proportionally using largest remainder. These rules were calibrated against 18,000 historical multi-member ward results, where the model achieves over 80% exact-match accuracy on seat allocation.
Accuracy
We backtest against every English local election from 2022 to 2025. Ward baselines and polling data are set to what was available before each election, though some model parameters (swing caps, conversion factors) were tuned across the full period rather than frozen at each point in time.
| Year | Councils | Seat error (RMSE) | Control correct |
|---|---|---|---|
| 2022 | 169 | 3.22 | 71% |
| 2023 | 230 | 3.42 | 71% |
| 2024 | 134 | 2.78 | 81% |
| 2025 | 219 | 4.92 | 66% |
| Overall | 3.58 | 72% |
RMSE is the average seat prediction error per party per council. An RMSE of 3.58 means the model is typically off by about 3-4 seats per party in an average council.
Control accuracy is how often we correctly predict which party (or none) runs each council.
2025 is notably worse. Reform went from near-zero to the largest local party by projected national share in a single cycle - a party jumping from 2% to 30% breaks any model built on historical ward baselines. The 2026 model benefits from having 2025 data as calibration.
The backtest covers 752 council-years across four election cycles. None of these are the exact set of 136 councils being predicted for 2026. London boroughs appear in 2022, metropolitan boroughs across 2022-2025 (by thirds), counties only in 2025.
Independents and Others are structurally underpredicted by 150-240 seats per year in backtest. This is not a swing or calibration error - it is a data gap. Many independent candidates do not appear in DCLEAPIL baseline data because they stood for the first time, or stood under a different label, or stood in a ward that has since been abolished. The model cannot predict candidates it has never seen. When April 2026 nominations close and we know exactly which independents are standing, this gap will narrow significantly. Until then, treat Others projections as a lower bound.
Known biases
| Party | 2024 error | 2025 error | Pattern |
|---|---|---|---|
| Conservative | -111 | -363 | Under-predicted |
| Labour | +251 | -9 | Over-predicted (2024) |
| Lib Dems | +64 | +460 | Over-predicted (candidacy) |
| Green | -1 | +140 | Over-predicted (candidacy) |
| Reform | -6 | +36 | Slight over-prediction |
| Others | -197 | -264 | Consistently under-predicted |
Lib Dems are the biggest overcount. The model over-predicted LD seats by 460 in 2025. Treat LD seat projections as an upper bound.
Independents and local groups are undercounted by 200+ seats in every year tested.
Reform was heavily undercounted in earlier model versions because ward baselines predated Reform's existence. Calibration against 2025 county results largely fixed this: the current model slightly over-predicts Reform by 36 seats in the 2025 backtest.
Conservatives are under-predicted, partly because incumbent councillors in shire and suburban wards outperform their national party more than the model captures.
Labour is over-predicted by a smaller margin, with the model awarding too many narrow ward wins.
Updates and candidacy
Predictions update automatically with every new published poll. The model runs on a rolling average of the seven most recent polls, so any new poll from a registered pollster shifts the numbers. We also publish a model changelog documenting every fix and improvement.
Two step-change improvements are coming:
April 2026: Democracy Club publishes actual candidate lists from the Statements of Persons Nominated. This will sharpen Reform and Green predictions in wards where they are not standing. It will not fix the bulk of the Lib Dem overshoot, which is structural rather than a candidacy problem.
Late April: A final forecast incorporating complete candidacy data and late polling.
Other forecasts
Most published local election forecasts use uniform national swing at the council level, converting vote shares to seats with a cube-law formula. That approach cannot account for geographic variation in swing, incumbency, or which parties actually contest each ward.
Published forecasts in 2025 were wrong by several hundred seats on multiple parties. Local elections are harder to forecast than generals, and no public model has a strong track record at council level.
We have not tested this model in a live election yet. 2026 is the first. The backtests look decent across 752 council-years, but the six-party fragmentation in 2026 is beyond anything in the training data.
Where it's most likely to be wrong
County councils. Reform is projected to win control of Essex, Norfolk, and Suffolk - our boldest calls. County ward baselines date from 2021 and predate Reform's local existence. Conservative seat projections in eastern counties reflect the combined effect of a halving in national Conservative support and strong Reform entry. In the 2025 backtest, the model under-predicted Reform in comparable councils (Lincolnshire: predicted 17, actual 44), suggesting these projections may if anything be conservative. The confidence intervals are wide for a reason.
Councils with strong independent traditions. 23 councils have more than 20% of seats projected for independents or local parties - Bradford, Oldham, South Tyneside, Bolton, Peterborough, the Isle of Wight among them. A single candidate standing or retiring can move 5-10 seats. These councils are flagged in the projector.
Candidacy gaps (pre-April). Until actual candidate lists arrive in April, we do not know which parties are standing in each ward. The 3% baseline proxy misses cases. Lib Dem and Green seat counts will adjust downward when real candidacy data is incorporated.
Two-candidate wards. In some metropolitan boroughs, wards with only two candidates in the baseline election produce inflated vote shares for the winning party. When a ward had only Labour and one other candidate, Labour's baseline can appear artificially high. South Tyneside is the clearest example, where several wards had two-candidate contests in 2022 giving Labour 60-70% baselines that overstate genuine support. This is a known limitation that will partially resolve when April candidacy data provides a wider field.
Inner London Green projections. Several inner London boroughs show large Green gains from low baselines - for example, Wandsworth Greens averaged 7% in 2022 but are projected at 11 seats. These projections assume Greens will contest every ward with strong candidates, and that national Green growth translates to wards where local Green organisation is thin. Boroughs like Hounslow (16% Green baseline, established ward organisation) are more robust. Councils where the Green projection depends heavily on candidacy assumptions are flagged in the data. These projections will be revised when April nomination data confirms which wards have Green candidates.
Confidence
Local elections have lower and more variable turnout than generals, patchy candidacy, and wards dominated by local factors. Seat-level predictions are noisier than for Westminster.
When we say a party has 800 seats with a range of 650-950, the range matters more than the point estimate. Council control probabilities - a 90% chance of staying Labour, say - are the most useful output.