Data quality, backtest

GE2024 backtest

Predicting GE2024 from the GE2019 baseline and the actual GE2024 polling, measuring whether our swing model adds value over Uniform National Swing.

Generated 2026-05-27 · 649 boundary-stable constituencies evaluated

Headline

Strong Transition Model vs Uniform National Swing

Real-polling forecast test (genuine pre-election simulation)

Uses ACTUAL May 2024 polling averages (Lab 44%, Con 24%, Reform 12%), what a forecaster running this model in May 2024 would have produced. The most honest forecaster comparison.

ModeWinner accuracyMajor-party MAEBrier
Uniform National Swing (May 2024 polling) 71.2% 5.57pp 0.3384
Strong Transition Model (May 2024 polling) 77.5% 5.06pp 0.3543

STM lift over UNS (real polling): 0.51pp. For comparison: Survation 2024 final MRP scored 83.8% / 5.55pp RMSE; YouGov 92% (in-campaign panel of 150k); UNS counterfactual 7.61pp RMSE.

Perfect-polling structural test (model only)

Feeds the model GE2024 actual national shares as if it had perfect foresight. Isolates the structural model quality from polling-error noise.

ModeWinner accuracyMajor-party MAEBrier
Uniform National Swing 67.3% 5.54pp 0.3812
Strong Transition Model 78.0% 4.52pp 0.3938
STM + anti-attenuation 71.2% 4.97pp 0.3749

STM lift over UNS (perfect polling): 1.02pp · Polling-error penalty: 0.54pp (extra MAE we paid for using May 2024 polling vs perfect foresight) · Anti-attenuation under-shoots even with out-of-sample calibration on GE2010+2015+2017 because pre-2017 distributions had wider between-PCON variance than 2024's fragmented 4-way race.

Per-party performance (STM)

Mean absolute error by party

PartyMAE
DUP 25.40pp
Labour 7.68pp
Reform UK 4.06pp
Liberal Democrats 3.97pp
UUP 3.60pp
Green Party 3.52pp
Conservative 3.35pp
Independent 2.70pp
Other 2.54pp
The Yorkshire Party 1.96pp
Alliance 1.33pp
Workers Party of Britain 1.28pp

Worst residuals

Where we missed

ConstituencyCountryMajor-party MAE
North Shropshire england 16.48pp
Bradford West england 14.66pp
Bristol Central england 14.65pp
Finchley and Golders Green england 13.39pp
Birmingham Hall Green and Moseley england 13.29pp
Doncaster North england 13.29pp
Bethnal Green and Stepney england 12.56pp
Manchester Rusholme england 12.49pp
Islington North england 12.40pp
Chelsea and Fulham england 12.33pp

Methodology · Forecast · Transparency