One model gives a party 280 seats and another gives it 330. The difference appears to settle whether it could govern alone, but the two totals can obscure overlapping fieldwork, different assumptions and dozens of close constituency calls.
A projected total is easier to understand when you can inspect the contests which produce it. Start there, rather than judging a model by whether its headline looks politically plausible.
Establish what is being estimated
Some models describe an election held under current voting intentions. Others attempt to forecast a future polling day, potentially accounting for campaign movement. Those are different tasks.
Identify the geography, parties, boundary set and treatment of seats outside the main survey. A GB model needs an explicit approach to Northern Ireland before its numbers can be compared with the 650-seat House of Commons.
Check fieldwork separately from publication. Two models released together may contain interviews from different political periods. An old estimate does not become current because it is republished on a new graphic.
Which method produced the numbers?
MRP models constituency preferences using survey and population information. A uniform swing calculation applies a national change to a previous local result. Other models combine approaches or add assumptions about local contests.
These methods can produce different patterns even when they begin with similar national shares. A comparison should identify the approach, input data and assumptions, rather than group every seat estimate under “polling”.
The British Polling Council’s MRP explanation provides background on the modelling process. Providers’ own release notes are needed for the particular implementation.
Close seats can amplify small differences
In an illustrative model, Party A leads narrowly in 40 constituencies. A small shift against A across those contests can remove many projected wins while changing its national share only modestly.
That sensitivity is a consequence of First Past the Post. A 0.2-point lead and a 20-point lead each produce one winner in a simple point-estimate map, although their evidential strength is very different.
Ask how many projected gains have comfortable margins and how many depend on small leads. Our guide to marginal seats explains why the runner-up and the wider field deserve attention.
A point total is not always an expected total
Counting the seat in each constituency where a party has the highest central share gives one kind of total. Adding that party’s win probabilities gives its expected number of seats. These need not match.
Across ten illustrative seats with an estimated 60% chance of victory in each, calling the most likely winner gives ten wins. The expected total is six. Neither description establishes the most likely joint national outcome, because that also depends on how the seat results are related.
Errors across constituencies can be correlated. A common national polling miss or a misjudged demographic relationship can move many seats together. Independently flipping each seat in a simulation can understate that shared risk.
Comparing ranges and claims
Read what a provider’s range means: a confidence or credible interval, simulation percentiles and an informal scenario range are not interchangeable. YouGov’s 2024 review is an example of an organisation examining both national and constituency performance after the result.
Do not average headline seat totals and call the result a probability model. An average of models can be a descriptive summary, but uncertainty needs a method of its own.
Finally, distinguish a majority from government formation. A hung Parliament can support several governing arrangements. A seat chart does not identify which agreements parties would make.
Sources and further reading
- British Polling Council: the MRP approach
- YouGov: assessment of its national and seat estimates
- UK Parliament: hung Parliaments
Reviewed 7 October 2026. Worked examples are illustrative unless an election or poll is identified.