Victorian election forecast
The forecast
How the model currently sees each seat. Before election day this is the pre-count forecast. From 6 pm on 28 November it updates as polling places report, and this page checks for a new version every minute.
What I did
A seat-by-seat forecast of Victoria's 88 lower-house districts, plus the 40-seat upper house, built so that the same models update live as results come in on election night.
- Booth-level swing model: results for every polling place in 2014, 2018 and 2022, mapped onto the current boundaries. Swings are split into statewide, regional, demographic (11 Census components, from a principal-components analysis of small-area data) and local parts, and that split sets how much a district can depart from the statewide polls.
- Swing form: a fitted mix of uniform and proportional swing (0.43 on a scale from 0 for uniform to 1 for proportional), estimated on past Victorian and South Australian elections. Proportional swings matter when a party's vote collapses: it falls furthest in its safest seats.
- Polls: a polling average sets each party's statewide vote and its uncertainty, drawn jointly so polling errors move together across parties. Each pollster's two-party figures are corrected by its measured error at the 2026 South Australian election.
- Preference flows: Victoria counts preferences in full, so who finishes in the final two and how preferences split decides most seats. The flows come from how-to-vote cards: a card-following model fitted on the full distributions of preferences in South Australia (2026) and Victoria (2022), which tracks every bundle of votes back to the candidate it started with.
- Simulation: thousands of simulated elections, each with its own statewide polling error, regional and local swings, and a full preference count in every district, give each seat's chances and the distribution of seat totals.
- Election night: as polling places report, each one updates the swing estimates for its region and district, with early, postal and absent votes modelled separately because they lean differently from votes cast on the day.
- Upper house: a full single-transferable-vote count engine for the eight five-member regions, tested against real counts from the 2025 Senate (Victoria) and the 2026 South Australian Legislative Council.
What I found
Replaying the 2022 election in the order results came in, with the model trained only on 2018:
| Counted so far | Seats called | Wrong calls | Margin error (points) |
|---|---|---|---|
| 10% of polling places | 72% | 1.5% | 6.0 |
| 40% of polling places | 80% | 0.5% | 4.3 |
| All polling places | 82% | 0% | 3.9 |
| Plus half the early votes | 92% | 0% | 2.0 |
Adding district-level swing priors did most of the work: with a uniform swing alone, about 11% of early calls were wrong.
The how-to-vote card model also improved how well preference flows could be predicted before any counting. Using only lodged cards and federal flows, it explained 77% of the variation in where preferences went at South Australia's 2026 election, against 69% from federal flows alone.
How the forecast did on 28 November: seats called, margin error, interval coverage
Limitation
The model's uncertainty is a little too narrow: in the 2022 replay, its 80% intervals for district margins contained the result about 72% of the time. It also has to borrow preference behaviour from other elections, so it is weakest where parties meet in contests that have rarely happened before.
- Methodology note: link needed
- Code: link needed