Methodology
Last updated: April 2026
Most UK election forecasters show you a single number. We show you the range of worlds you might be living in. Here is how the model works — and what it cannot do.
General election model
Coverage
The model covers 632 constituenciesin England, Scotland and Wales. Northern Ireland's 18 seats are excluded as they have a separate party system (DUP, Sinn Féin, Alliance, SDLP, UUP) not captured by GB-wide polling. All seat projections, probabilities and confidence intervals refer to this 632-seat GB parliament.
Poll aggregation
National polling data is ingested from all BPC-registered pollsters via an automated scraper running every 6 hours against Wikipedia's UK polling tables. Currently over 500 polls from July 2024 onwards.
Polls are weighted by three factors. Recency:exponential decay with a 10-day half-life for the current snapshot — recent polls dominate, and polls older than 3 weeks contribute very little. Sample size: larger polls get more weight. Pollster accuracy: house effect corrections based on each pollster's systematic bias vs the 2024 general election result.
Seven-layer seat projection pipeline
Simple poll-averaging tells you national vote shares. Uniform national swing converts those into seats. Both approaches assume the country moves as one block. It doesn’t.
Our model asks a harder question: who turns up, and where do the votes come from? The same national polling could produce anywhere from 27 to 243 Reform seats depending on turnout patterns and tactical voting dynamics. That uncertainty is real, and we show it honestly.
The model pipeline has seven layers. Each one transforms the raw projection and adds a different kind of intelligence.
Regional swing
National vote shares are disaggregated into regional projections. Scotland, London, the North, and the South behave as distinct political systems. Direct polling where available (Scottish polls, London-specific data) replaces national uniform swing. Regions without direct polling get wider uncertainty bounds.
Reform vote source model
Reform’s vote share doesn’t come equally from all parties. Using BES voter flow data, the model estimates differential switching rates: Conservative-to-Reform is much higher in Leave-voting constituencies, Labour-to-Reform is concentrated in working-class northern seats. This produces geographic variation in Reform’s projected strength that uniform swing cannot capture.
Green vote source model
Same approach for the Green surge — Green gains are concentrated in university towns, inner cities, and seats with large young and progressive populations. The geographic pattern is very different from Reform’s.
Correlated turnout scenarios
Both Reform and Green are drawing support from non-traditional voters whose turnout at a general election is uncertain. If non-traditional voters mobilise, both insurgent parties benefit simultaneously — this is a correlated risk. The model runs three turnout scenarios (established voter dominance, baseline, mobilisation wave). The difference between scenarios spans over 200 seats.
Tactical voting
In FPTP marginals, some voters will vote strategically rather than sincerely. The model applies an 18% mean squeeze rate — selectively by seat type. Full squeeze in clear two-horse races, half in seats with a competitive third place, quarter in genuine three-way contests. This was calibrated after finding that a 30% squeeze rate produced implausible results.
Constituency classification
Each of the 632 GB seats is classified by its competitive structure (Lab-Ref, Lab-Con, Con-LD, three-way, safe, etc.). 523 of 632 GB seats changed competitive structure between 2019 and 2024, reflecting the scale of political realignment. The model applies different swing dynamics to different seat types.
Monte Carlo simulation
1,000 simulated elections drawing from all probability distributions simultaneously: polling error, turnout scenarios, tactical voting rates, and regional variation. Each simulation applies correlated noise — if Labour overperforms nationally, they tend to overperform in Labour-leaning marginals. The output is a full probability distribution: P(Labour majority), P(Hung parliament), P(Reform largest party), and seat-level win probabilities.
Model Waterfall
How each model layer transforms the raw projection. Radical transparency — see exactly what drives the forecast.
Backtest results
The model was backtested against the 2024 general election using only polling data available before July 4, 2024:
The 4 errors were all independent or non-party winners that no national model can predict. Phase 1 of the model (simple uniform national swing) achieved 552/650 (84.9%). The seven-layer pipeline improved accuracy by 94 seats.
Current projection
As of the latest model run (Monte Carlo means with 90% confidence intervals):
Outcome probabilities: P(Labour majority) 69.2%, P(Hung parliament) 30.8%, P(Reform largest party) 6.5%.
Why turnout matters more than polls
In 2024, Reform polled at 15–17% in final polls but got 14.3% on the day. Their supporters didn’t turn out at the rate the polls implied. This is the established pattern: non-traditional voter bases underperform on election day.
But 2016 showed the opposite. Brexit referendum turnout was exceptionally high in demographics that usually stay home. Non-traditional voters canmobilise — you just cannot predict when from polling alone.
Our three turnout scenarios capture this uncertainty honestly. The difference between “established voter dominance” and “mobilisation wave” is the difference between Labour winning comfortably and Reform becoming the largest party.
Key uncertainties and limitations
Reform's ceiling. The model's widest confidence interval is Reform at 27–243 seats. The range is this wide because Reform has never contested a general election at this polling level, and the geographic efficiency of their vote is genuinely unknown. Small changes in how their ~27% national vote is distributed across constituencies produce enormous swings in seats — this is the defining feature of FPTP with five competitive parties.
Liberal Democrat seat range. The model projects the Liberal Democrats at 80 seats with a confidence interval of 77–82 — a range of just 5 seats. This is the narrowest interval of any party and may understate the true uncertainty. The 2024 election saw the Liberal Democrats surge from 11 to 72 seats through an exceptionally efficient tactical voting campaign targeting Conservative-held southern seats. Many of these are first-term incumbencies won on thin margins during the anti-Conservative wave. The model treats them as secure, reflecting the strong historical pattern that Liberal Democrat incumbents outperform their national polling.
However, the political landscape has shifted significantly. With Reform replacing the Conservatives as the dominant right-of-centre force in many areas, the tactical logic that delivered these seats — “vote Liberal Democrat to remove the Conservative” — may not hold in the same way. Whether the Conservative-to-Reform switch helps the Liberal Democrats (by permanently splitting the right) or hurts them (by creating unpredictable multi-party contests where the tactical coalition fractures) is genuinely uncertain. A more realistic range might be 60–95 seats, and we expect to revisit this after calibration with real election data.
Tactical voting dynamics. The 18% squeeze rate was calibrated against 2024 patterns, when tactical voting was predominantly anti-Conservative. In a five-party landscape where the tactical logic is more complex (anti-Reform? anti-Labour? pro-local?), the direction and magnitude of tactical voting is harder to predict. The model's tactical layer is the most assumption-dependent component.
Independent candidates. The model has no way to predict when a local independent will outperform. Our 2024 backtest gets 646/650 seats right — the 4 errors are all independent wins driven by hyperlocal dynamics invisible to national polling.
Campaign effects and black swan events. A brilliant or terrible campaign in a specific constituency can swing the result. A major scandal, economic shock, or leader change would invalidate the current polling inputs. The model updates automatically when new polls arrive, but it cannot predict discontinuities.
The next election is years away. Current projections assume an election held today. The actual general election is not required until 2029. Political conditions, party leadership, economic fundamentals, and voter preferences will change substantially before then. The forecast's value is in mapping the current landscape and identifying structural dynamics — not in predicting an election that is years away.
Data sources
- Polling data: All BPC-registered pollsters via automated Wikipedia scraper (500+ polls, updated every 6 hours)
- 2024 general election results: Electoral Commission, all 650 constituencies on 2024 boundaries
- Voter flow estimates: British Election Study Internet Panel
- Constituency demographics: Census 2021, used for regional swing calibration and vote source models
- Constituency boundaries: ONS GeoJSON for 2024 parliamentary boundaries
- Scottish and Welsh polls: Direct regional polling, ingested on the same 6-hour cycle
- Prediction markets: Betfair exchange odds, ingested daily for cross-reference
Connection to local elections model
The GE forecasting model provides the geographic anchor for the local elections projection — particularly for Reform UK, which had near-zero local presence before 2025. Reform's constituency-level GE projections are mapped to council ward boundaries to estimate their local vote share. This means the two models are structurally linked: changes in national GE polling propagate through to local election projections.
After the May 7, 2026 local elections, the reverse flow will be equally important. Actual local election results will provide calibration data for the GE model's parameters — particularly the Reform local discount factor, the protest vote adjustment, and the tactical voting multiplier. We will publish a post-election calibration analysis and update the GE forecast accordingly.
Local elections model
Last updated: April 2026. The local elections model is distinct from the general election model. Different inputs, different mechanics, a different uncertainty profile.
Vote source model
Most local election forecasts apply uniform national swing — if Labour is down 15 points nationally, every ward loses 15 points. This breaks when a party goes from near-zero to 25% in four years. Our model instead tracks where blocks of voters are moving. Using vote flow rates from the British Election Study Internet Panel — which follows 30,000 voters over time and records exactly which parties they switch between — the model estimates how many 2022 Conservative voters now support Reform, how many 2022 Labour voters have moved to the Greens, and how many have dropped out entirely.
The model applies these flows differently across five council profiles (London boroughs, metropolitan boroughs, county councils, unitary authorities, and shire districts), each with BES-calibrated vote flow rates reflecting distinct local political environments.
Ward-level geographic distribution
Within each council, party vote shares are not applied uniformly. Where 2022 ward-level results exist (covering roughly 75% of wards), the model uses the historical ward-level pattern as a distribution template. A ward where the Greens polled 40% in 2022 gets a larger projected Green share than one where they polled 5%, even when the council-wide average is the same. This preserves the geographic concentration that is critical to seat outcomes under first-past-the-post.
The concentration effect is dampened at 70% rather than fully preserved, allowing for some expansion of support beyond 2022 strongholds while preventing the model from being imprisoned by four-year-old data. Wards without 2022 baselines receive a discounted estimate based on the council-level projection and, where available, 2024 general election constituency results mapped to ward boundaries.
Reform UK: the geographic anchor
Reform contested almost no local wards before 2025. With no 2022 ward-level baseline to distribute from, the model uses general election constituency-level projections as a geographic anchor, mapped to council wards via ward-constituency boundary overlaps. A local discount factor (0.55–0.80 depending on scenario) accounts for the pattern that insurgent party support is lower in local elections than general elections.
Reform's local conversion rate is the single largest source of uncertainty. Reform won 41% of seats with 31% of votes in 2025 — a seats-to-votes ratio previously only achieved by established parties with deep local infrastructure. Whether this was a one-off breakthrough or a new normal is genuinely unknown. The three scenarios reflect this uncertainty directly.
Multi-member ward allocation
London boroughs and all-up metropolitan boroughs elect three councillors per ward simultaneously. The dominant pattern is slate voting — people vote for all three candidates from one party. This creates winner-takes-all amplification: the party with the plurality typically wins all three seats.
The model replicates this using margin-based seat allocation calibrated against historical multi-member ward results. Where the leading party has a comfortable margin, it sweeps all seats. In close two-party races, seats split two-to-one. This amplification means Labour's inner London strongholds deliver far more seats per ward than a proportional model suggests, and Reform's northern victories are more decisive.
Monte Carlo simulation and confidence
The model runs 1,000 simulated elections per scenario, varying ward-level vote shares with random noise calibrated to historical election variance. This produces probability distributions rather than point estimates. Council control outcomes use the fraction of simulations producing each result — outright majority, largest party, or no overall control.
NOC probabilities are systematically higher than in naive single-point forecasts, because the simulations correctly capture that small vote share changes can flip councils from majority to minority. In councils where the top two parties are within four points, no overall control is frequently the modal outcome.
Coverage and data sources
The model covers all 136 English councils holding elections on 7 May 2026: 32 London boroughs, 32 metropolitan boroughs, 18 unitary authorities, 6 county councils, and 48 district councils, representing 5,004 contested seats. Projections incorporate confirmed candidate lists from Democracy Club's Statements of Persons Nominated (9 April 2026).
Data sources: national polling from BPC-registered pollsters (6-hour update cycle); BES Internet Panel vote flow rates (May 2025 wave); 2022 ward-level results from Open Council Data; GE constituency projections from the Reading Signal 7-layer pipeline; council reference data from Rallings and Thrasher (Exeter University); external calibration from Stephen Fisher (Elections Etc) and PollCheck.