How the forecasts work
Methodology
Every forecast on The Margin runs on the same engine. It reads the published polls, starts each seat from its last result, plays the election 50,000 times with realistic luck in every run, and counts the preferences in every seat. This page explains each step, for Victoria, NSW and the federal election alike.
The short version
- Average the polls. Recent, larger and better-documented polls count for more. Each pollster's habitual lean is estimated and removed, and the latest polls are steadied by the average of the past year of polling.
- Simulate the vote. Each simulation draws a statewide (or national) result around that average, with room for a polling miss and for opinion to move before election day.
- Move every seat. Each seat starts from its last result and swings with the simulated vote, plus regional, neighbouring-seat and local swings, so seats move together the way real seats do.
- Count preferences. In every seat of every simulation, candidates are eliminated one by one and their preferences passed on until two are left.
- Fill the upper house. Each simulated vote also fills the Legislative Council or the Senate with a quota count.
- Count the outcomes. A probability is the share of the 50,000 simulations in which something happens. “6 in 10” means it happened in 30,000 of them.
Odds shown on the forecasts are fair odds (1 ÷ the chance), with no bookmaker margin. They are another way of reading a probability, not betting advice.
What goes in
- Past results. First preferences and full distributions of preferences from the Victorian Electoral Commission, the NSW Electoral Commission and the Australian Electoral Commission, on the boundaries that apply at the coming election.
- Boundaries. Official electorate boundaries, used to find each seat's neighbours and to draw the maps.
- Polls. Published voting-intention polls, including state and regional breakdowns where pollsters report them, and single-seat polls reported in the media.
- Candidates. Who is standing where. Until nominations close, each party's chance of standing in each seat is itself simulated.
- Economic and political conditions, for the polls + context model only (see Context).
Every input is frozen into a dated, fingerprinted snapshot, so any published forecast can be rebuilt exactly.
From polls to a statewide vote
The forecast's view of the overall vote comes from polls alone. Past election results decide where each party's vote falls seat by seat, but they do not pull the overall forecast.
Which polls count
A poll needs a sample size, fieldwork dates and a published question so it can be checked. Polls missing any of these are listed with the reason, but left out of the average. Polls more than a year old are ignored, and a poll that leaves out a major party cannot be compared with the rest, so it is left out of the headline too.
How much each poll counts
- Recency. A poll's weight fades smoothly with age, and fades faster in the last months of a campaign, when opinion is settling.
- Disclosure. Pollsters that meet the Australian Polling Council's full disclosure standard count most; polls with only partial methods count much less and are flagged.
- Who paid. A poll commissioned by a party or an interest group is discounted unless it is released in full.
- Sample size enters only through sampling error. Every poll also carries a floor of non-sampling error, so one very large poll cannot swamp the rest.
- One pollster is one pollster. That error floor is shared across all of a firm's polls, so ten polls from one firm do not count as ten independent views.
- House effects. Once a pollster has several usable polls, its consistent lean against the other pollsters is estimated, cautiously, and removed before averaging.
Latest polls against the past year
The average blends the latest polls with a long-memory average of the past year of polling. The more good, recent polling there is, and the closer the election, the more the latest polls count; a single unusual poll cannot swing the forecast. A gap in polling does not erase the last polls: the time since then is treated as extra uncertainty. Past election results are used only if there is no polling at all.
Where the vote sits
When a poll also reports a region or state (regional NSW against Sydney, Melbourne against regional Victoria, or each state in a national poll), that figure is compared with the same poll's overall number, and each party's vote is shifted between regions to match, part-way, depending on how many voters in that region were asked. The overall totals do not change.
Uncertainty
Every simulation adds sampling error, pollster error, movement before election day and an overall polling miss shared by every seat. Errors are drawn so that vote shares always stay positive and add to 100%. A party far from its last election result, as One Nation is now, gets wider error in both directions: in past surges like this, final polls have missed by several points either way, so the model widens the range rather than pulling the forecast back towards the old result.
From the overall vote to every seat
- Start from the last result. Each seat begins from its last primary vote on the boundaries that apply now. Where boundaries have changed (the new South Australian, Tasmanian and ACT federal divisions), it begins from an estimate of the last result on the new boundaries, built from enrolment transfers and booth results.
- Swing proportionally. Each party's vote moves in proportion to its local strength, calibrated in every simulation so that the seats add up to the simulated overall vote. A party that doubles its vote gains most where it was already strong.
- New candidates. Where a party had no candidate last time, the seat starts from that party's upper-house vote in the same area, which every voter cast, or failing that from its strength in similar seats nearby.
- The candidates themselves. Sitting members carry a personal vote that is lost when they retire; first-term members tend to gain. Independents keep their own local baselines rather than following a statewide swing.
- Seats move together. On top of the overall swing, each simulation draws a regional swing, a shared swing between neighbouring seats and a local swing for each seat, separately for each party. A good night for one party in western Sydney therefore shows up across western Sydney, not in one seat at random.
- Seat polls. A single-seat poll is blended with the model's own estimate, with modest weight: seat polls are small, often commissioned, and historically less accurate than statewide polls. Polls that are poorly documented count less still, and a poll in one seat also nudges similar seats nearby.
Counting preferences
Every seat in every simulation is counted the way the electoral commission counts it: the last-placed candidate is excluded and their ballots passed on, until two remain. The final two can be any pair, including One Nation, Greens or independents.
Where preferences go
Preference flows come from the distributions of preferences at past elections. Each seat's flows are estimated from its own count where it has one, pulled towards its region's and the whole state's flows where its own evidence is thin. Each simulation draws its own flows around those estimates, so preference uncertainty is part of every forecast.
Optional and compulsory preferences
- NSW uses optional preferential voting: voters can stop numbering, so some ballots exhaust and play no further part. The model counts exhaustion as it happens in NSW.
- Victoria and the federal House use full preferential voting: every box must be numbered, so no vote exhausts.
Preferences to One Nation
At recent elections, Liberal and National candidates were almost never excluded while One Nation was still in the count, so past counts record very little about how Coalition voters rank One Nation. Left alone, that gap would send Coalition preferences almost entirely to Labor in a Labor against One Nation contest. Instead, the model estimates how strongly each party's voters favour each other party from every exclusion in the 2025 federal count, with Liberal, National, LNP and CLP voters pooled. This choice matters a great deal wherever Labor and One Nation are close, and it is applied the same way in all three forecasts.
The upper house
Each simulation's vote also fills the Legislative Council or the Senate. Each party's upper-house vote follows its simulated lower-house vote, steadied by its past upper-house results and any upper-house polling.
- NSW: 21 of the 42 Legislative Council seats are elected statewide; the other 21 members continue. A group needs about 4.5% of the vote for a quota.
- Victoria: all 40 seats, five in each of eight regions. Group voting tickets are gone, so preferences follow the voters, and are taken from Assembly preference flows with a share of each transfer exhausting.
- Senate: 40 of the 76 seats, six in each state and two in each territory, each its own quota contest, with transfers from the 2025 Senate count. The 36 senators elected in 2025 continue.
This is a group-level quota approximation, not a full replica of the official count: it allocates seats by quotas and passes preferences between groups, rather than replaying every ballot.
Polls only, or polls + context
Every forecast comes in two versions. Polls only uses everything above. Polls + context adds a small amount of information about the economy and, for the two states, the federal political climate.
Voters judge governments partly on how life is going: studies of nearly two hundred Australian state elections, and of federal elections, find that economic conditions shift a government's chances by modest amounts. State results have also tended to run against whichever party holds office in Canberra. In a federal election there is no separate federal government to punish, so the federal forecast's context is the economy alone.
The same research is clear that these effects are small and noisy, so the adjustment is:
- Small and capped. It can nudge a forecast but never overturn it.
- Uncertain by design. Each simulation draws its own strength for the effect, so context widens the range of outcomes as well as shifting it.
- A prior, not a poll. It adjusts the starting expectation, never the polls themselves.
- Fading. As good polling accumulates, polls take over and context fades towards zero.
Both versions run the same 50,000 simulations with the same random numbers, so any difference between them comes from context, not chance. Each forecast shows how much context changes it.
The specific indicators and how much each counts are not published.
The three elections
| Victoria | NSW | Federal | |
|---|---|---|---|
| Election | 28 November 2026 | 13 March 2027 | Not yet called; assumed 20 May 2028, the latest possible joint date |
| Lower house | 88 districts; 45 for a majority | 93 districts; 47 for a majority | 150 divisions; 76 for a majority |
| Starting point | 2022 results; boundaries unchanged | 2023 results on current boundaries | 2025 results; estimated results on the new SA, Tasmanian and ACT boundaries |
| Regions for shared swings | The eight Legislative Council regions | Ten regions, from western Sydney to the western rural seats | Twelve: capital and rest of state for the big states, plus SA, Tasmania, ACT and NT |
| Preferences | Full (compulsory) | Optional | Full (compulsory) |
| Upper house | All 40 Council seats, 8 regions × 5 | 21 of 42 Council seats, statewide | 40 of 76 Senate seats |
| Context | Economy and federal climate | Economy and federal climate | Economy only |
“Government” means a party winning a majority on its own; the Liberals and Nationals (and the LNP and CLP) count as one Coalition. “No majority” means none of Labor, the Coalition or One Nation reaches a majority in that simulation. The tipping-point seat is the one that most often delivers a party's majority-making seat in the simulations where it wins one.
What it can't tell you
- Not yet backtested. The weights are reasoned defaults, not yet tuned against past elections. Treat the ranges as honest guides rather than calibrated guarantees.
- One Nation is hard. A party polling far above its last result has little local history to go on, so its seat-by-seat numbers are the least certain on the site.
- Before nominations close, who stands where is simulated, not known.
- The federal date is assumed, and Queensland's federal seats stay on 2025 boundaries until its redistribution is final.
- Upper houses are approximated at group level, as described above.
Polls are noisy and every model is a simplification. A 25% chance happens one time in four.