winners called correctly across 2,977 wards in the May 2024 local elections
Accuracy record
How accurate are these forecasts?
Before any forecast on this site is published, the same model is run against the most recent real election to see how well it would have done. The full results live here, including every category where it was wrong.
Headline numbers
How well the model would have called past elections
average gap between predicted and actual party share, May 2024 local backtest
winners called correctly across 649 constituencies in the 2024 general election
average gap between predicted and actual party share, 2024 general election backtest
Honesty rules
Things we won't do
- We don't fit on the test set. The 2024 backtests use only what a forecaster running the model in May 2024 would have known, no peeking at the actual results during model fit.
- We don't quietly retire models that did badly. The same model code that generated the backtest also generates the live forecast. If we change the model, the backtest re-runs automatically and the new score is published, even if it's worse.
- We publish the worst misses, not just the headline. Every backtest page lists the wards or constituencies where the model was furthest off, with no cherry-picking.
- We pre-register the live forecast before each election. The sha256 of every prediction file is recorded ahead of time so we can prove there's been no after-the-fact tampering. See Transparency.
For comparison
How we sit against published forecasters
| Forecaster | Method | Winner accuracy on GE2024 |
|---|---|---|
| YouGov 2024 (fared paper) | 150k in-campaign panel; full Bayesian MRP | ~92% |
| Stonehaven 2024 | Bayesian MRP | ~94% |
| Focaldata 2024 | MRP with regradient | ~87% |
| Survation 2024 | MRP, no unwinding | ~84% |
| UK Elections (this site) | STM swing + per-constituency BES priors + tactical voting + demographic floors and ceilings, no panel fit | 78.0% |
| Uniform National Swing baseline | The simplest non-trivial model | ~71.8% |
The gap to Survation/Focaldata is essentially the in-campaign panel-fitted MRP step that fresh fieldwork enables. We don't run a fieldwork panel, so we won't catch up to that fully, but we're 5 percentage points ahead of Uniform National Swing, the standard baseline for forecasters without a panel.