Estimates

Fielding 300 interviews in each of 543 seats is expensive and still fragile. Multilevel regression with post-stratification does better: fit a model of vote choice across the whole national sample, then re-weight its predictions onto each constituency's own demographic composition. A seat with 74 interviews borrows strength from the 400 seats that resemble it. This is how a national broadcaster gets a seat map on a Tuesday instead of a state map in three weeks, and the post-stratification frame it needs is months of work that a rival cannot buy.

The second half is why 2024 broke. Vote share estimates were not badly off; seat forecasts were, because seat-level errors were treated as independent. They are not. If a poll is off two points nationally it is off two points nearly everywhere at once and hundreds of marginals flip together. Adding a shared swing term widens the interval roughly fourfold and, critically, contains the truth.

Which forces a presentational change. A broadcaster wants "NDA 361". The defensible output is a probability of a majority. Whether he can sell that to a television partner is the real commercial question on this screen.