How we forecast 650 constituencies

1. GE2024 baseline. Each constituency's actual result on 4 July 2024, taken from Democracy Club CC0 data + Returning Officer declarations. 650/650 constituencies.
2. National swing. Strong Transition Model (STM). Replaces additive Uniform National Swing. Losers shed share locally in proportion to their local vote; gainers absorb the freed share pro-rata to national gain. Bounded in [0,1] by construction. Reference: Electoral Calculus / Baxter STM. +1.0pp MAE lift over UNS on the GE2024 backtest.
3. BES regional/demographic prior. 36,416 post-GE2024 respondents from the British Election Study Wave 1-30 panel, post-stratified into (region × age × ethnicity × tenure) cells with Dirichlet pooling κ=15, then projected to LAD-level priors. Each constituency inherits the prior of its primary intersecting LAD, blended at 5% weight.
4. Sitting-MP layer. Personal-vote bonus for incumbents standing again (+1pp at <5yr → +4pp at ≥20yr tenure); retirement drag for standing-down MPs (-2pp to -3pp); open-seat treatment for defectors / suspended MPs.
5. Tactical voting overlay. Curtice/Fisher Lab/LD/Green progressive squeeze. When 1st = Conservative or Reform with ≤10pp lead over a progressive 2nd, transfer 15% of a 3rd-placed progressive's share to the 2nd-placed.
6. English-identity floor (Census 2021 TS027). Reform UK gets a small lift in constituencies where the share of residents giving an "English only" or "English and British" national identity is high. +1pp at 22%+, +2pp at 30%+, +3pp at 40%+. Frontiers (2025) "drivers of Reform UK support" finds English national identity is the single strongest demographic predictor of Reform vote (standardised β ≈ 0.45).
7. Age-structure adjustment (Census 2021 TS007A). Reform and Conservative get a small lift in constituencies with a high 65+ share. Dampened to 40% strength when a BES prior is already in play, since BES respondents implicitly carry the age signal.
8. Reform Muslim-share ceiling. Reform's predicted share is capped at the demographically-plausible ceiling (12% if the constituency is >50% Muslim, 22% if 30–50%, 30% if 15–30%). Excess is redistributed pro-rata. Calibrated against Sobolewska & Ford (Brexitland, 2020).
9. Independent ceiling. Independent vote capped at 8% in regular contests. The cap is bypassed for known high-profile independents (Galloway, Corbyn, Lloyd etc.).
10. By-election overlay. Where a post-GE2024 by-election has been held in this constituency, blend 30% of the by-election's shares into the prediction.
11. Normalise. Re-scale shares to sum to 1.0.

Predicting GE2024 from the GE2019 baseline + actual GE2024 national polling, evaluated on 649 boundary-stable PCONs:

• Uniform National Swing baseline: 67.3% winner accuracy, 5.54pp major-party MAE.
• Strong Transition Model: 78.0% winner accuracy, 4.52pp major-party MAE , a 1.02pp lift over UNS.

For comparison: Survation's published 2024 MRP RMSE was 5.55pp; YouGov's 2024 fared paper reported 92% winner accuracy. Our 78% on the boundary-stable 439-PCON subset is the floor, adding the new-boundary 211 PCONs once notional 2019 results are ingested, plus campaign-period polling, should close the gap.

Full backtest tables

Every weight, ceiling, and dampening factor is recorded in data/predictions/ge-next/assumptions.json and pre-registered with its sha256 before the forecast is run. Current values:

{
  "use_strong_transition_model": true,
  "ge_dampening": 1,
  "bes_prior_weight": 0.15,
  "tactical_voting_competitiveness_gap": 0.1,
  "tactical_voting_floor": 0.05,
  "tactical_voting_transfer_rate": 0.3,
  "by_election_overlay_weight": 0.3,
  "incumbency_personal_vote_pp": {
    ">=20yr": 4,
    ">=10yr": 3,
    ">=5yr": 2,
    "<5yr": 1
  },
  "retirement_drag_pp": {
    ">=20yr": 3,
    ">=10yr": 2.5,
    "<10yr": 2
  },
  "defection_open_seat_drag_pp": 1.5
}

• Lauderdale, Bailey, Blumenau, Rivers (2020), foundational MRP paper
• Hanretty, Lauderdale, Vivyan (2018), comparison of MRP strategies
• YouGov 2024 fared paper, unwinding correction
• Survation/Hanretty 2024 MRP post-mortem
• Focaldata 2024 review
• Electoral Calculus Strong Transition Model
• Fisher (2004), tactical voting and tactical non-voting

National forecast · Transparency & pre-registration · Backtest details