A poll can identify the winning party and still miss its vote share badly. A seat model can call most constituencies correctly while performing poorly in the handful of close contests that determine whether a government has a majority.
Accuracy needs a stated measure and a fair comparison. “Got the election right” is too loose a verdict for assessing either an individual organisation or polling as a whole.
Compare like with like
National voting-intention estimates should be compared with the election result for the same geographical coverage. A Great Britain poll excludes Northern Ireland; an unadjusted UK result includes it. Party definitions and the treatment of independents also need to match.
Timing matters. A poll conducted several weeks before voting is a measurement from that period, not necessarily a forecast of polling day. Final campaign polls face a shorter gap, but late decisions, postal voting and different fieldwork schedules can still complicate comparisons.
Our introductory polling guide separates what a poll measures from claims about a future result. Historical comparisons explains the equivalent problem with election data.
Different scores tell different stories
Mean absolute error averages the absolute gap between estimated and actual shares for a stated set of parties. In an illustrative comparison with errors of +2, −1 and −3 points, the mean absolute error is two points.
A signed average can hide that error: the positive and negative deviations partly cancel. Root mean squared error gives larger individual misses more influence. Neither statistic can be interpreted properly without knowing which parties were included.
A party lead is another measure. An estimate of 32% against 28% has a four-point lead; a result of 30% against 29% has a one-point lead. The lead error is three points. That does not mean every party share was three points wrong.
Look at several elections and the current method
A strong performance at one election is evidence, but it does not guarantee the next. Recruitment, political competition and turnout behaviour can change. So can the organisation’s own method.
The independent inquiry into the 2015 polling miss identified unrepresentative samples as its principal cause. That finding is more informative than a general claim that voters lied or that all polls are worthless.
The YouGov review of 2024 examined its different approaches and the effects of turnout and late movement. It is a pollster’s own review, which should be read as such, alongside other evidence. An honest account of weaknesses is useful even when a method performed comparatively well.
Seat accuracy needs separate assessment
In an illustrative election with 650 seats, predicting the winner correctly in 600 gives a 92.3% success rate. That headline can coexist with a large error in one party’s seat total if most wrong calls lean in the same direction.
The task also contains many easier calls in constituencies with large margins. An evaluation should examine performance in close seats, party totals and the uncertainty attached to forecasts, rather than stop at an overall percentage of winners identified.
What can be judged between elections?
Without an election result, the exact national voting intention on a past date is not observed. Disagreement with the average is a house-effect question, not proof of inaccuracy.
Between elections, examine disclosure, consistency, method changes and the breadth of evidence. Use our pollster directory to find organisations’ original material, and judge large claims against what that material actually says.
Sources and further reading
- BPC/MRS independent inquiry into the 2015 polls
- YouGov: review of 2024 performance
- House of Commons Library: verified 2024 results
Reviewed 7 October 2026. Worked examples are illustrative unless an election or poll is identified.