Backtest
How accurate was the model on the last cycle?
We took the same model that produces the 2026 forecast, gave it only data available before May 2024, and asked it to predict the actual May 2024 ward results. Here's how it did.
Headline
Two numbers
Per-party
Mean Absolute Error by party
| Party | MAE | Residual SD |
|---|---|---|
| Labour | 7.23pp | 11.11pp |
| Conservative | 5.37pp | 7.57pp |
| Reform UK | 6.41pp | 5.31pp |
| Liberal Democrats | 4.81pp | 6.84pp |
| Green Party | 4.64pp | 6.75pp |
MAE is the absolute difference between predicted and actual party share, averaged across all wards. Residual SD is used to calibrate the prediction intervals on individual ward pages.
Worst misses
30 wards where the model was furthest off
| Council | Ward | MAE | Predicted | Actual |
|---|---|---|---|---|
| pendle | Brierfield East & Clover Hill | 68.0pp | Labour | Independent ← |
| rochford | Lodge | 56.5pp | Conservative | Liberal Democrats ← |
| oldham | Werneth | 52.8pp | Labour | Independent ← |
| pendle | Bradley | 51.8pp | Conservative | Independent ← |
| rochford | Sweyne Park & Grange | 48.8pp | Liberal Democrats | Independent ← |
| blackburn-with-darwen | Shear Brow & Corporation Park | 46.0pp | Labour | Independent ← |
| burnley | Queensgate | 45.8pp | Labour | Independent ← |
| blackburn-with-darwen | Little Harwood & Whitebirk | 45.7pp | Labour | Independent ← |
| pendle | Marsden & Southfield | 45.0pp | Labour | Independent ← |
| pendle | Whitefield & Walverden | 41.9pp | Labour | Independent ← |
| blackburn-with-darwen | Billinge & Beardwood | 40.4pp | Labour | Independent ← |
| burnley | Bank Hall | 36.3pp | Labour | Independent ← |
| blackburn-with-darwen | Bastwell & Daisyfield | 35.8pp | Labour | Independent ← |
| burnley | Daneshouse with Stoneyholme | 35.1pp | Labour | Independent ← |
| tameside | St Peter's | 33.1pp | Labour | Independent ← |
| blackburn-with-darwen | Blackburn Central | 32.7pp | Labour | Independent ← |
| blackburn-with-darwen | Roe Lee | 30.0pp | Labour | Independent ← |
| peterborough | Stanground South | 27.9pp | Conservative | Peterborough First ← |
| burnley | Lanehead | 27.8pp | Labour | Labour |
| blackburn-with-darwen | Audley & Queen's Park | 26.4pp | Labour | Independent ← |
| rochdale | Central Rochdale | 26.3pp | Labour | Workers Party of Britain ← |
| west-lancashire | Rural South | 26.3pp | Independent | Our West Lancashire ← |
| kirklees | Batley East | 26.2pp | Labour | Independent ← |
| calderdale | Park | 26.0pp | Labour | Workers Party of Britain ← |
| south-tyneside | Hebburn North | 25.5pp | Labour | Labour |
| bradford | Toller | 25.4pp | Labour | Independent ← |
| bradford | City | 25.4pp | Labour | Independent ← |
| blackburn-with-darwen | Wensley Fold | 25.1pp | Labour | Independent ← |
| rochdale | Milkstone & Deeplish | 24.9pp | Labour | Workers Party of Britain ← |
| bradford | Little Horton | 24.0pp | Labour | Independent ← |
Most of the worst misses are wards dominated by strong local independents or single-issue parties (Our West Lancashire, Peterborough First, Tunbridge Wells Alliance, etc.). The model has no national-tier baseline for these and underestimates their pull. This is a known limitation; we surface it rather than hide it.
Why publish this?
Every forecast site claims accuracy. Few publish their errors. The 5.69pp average share error and 82.7% winner accuracy on May 2024 are the credibility floor . we'll publish the same backtest against May 2026 actuals after the count, so you can see whether the model improves or degrades over time.
Methodology: the model was given only ward history before May 2024, then asked to predict each 2024 ward result using May 2024 polling vs the GE 2019 baseline, plus local-authority demographic projections and Index of Multiple Deprivation deciles. See the methodology page for the full pipeline.