Simulating Australian polling error: where the chances come from

Summary

The seat ranges, the chance of a hung parliament, the chance that any party or combination reaches 76 seats and the chance in each division all come from 2,000 simulated elections. Each run moves every group's national primary vote by a random error the size of its past final-poll miss (1.25 to 1.87 points), adds a random state error of 1.5 points per group in every state, a random candidate error that grows with the candidate's share, and a 2.5-point error on the 2PP in divisions that end as Labor against the Coalition, then replays the count in all 150 divisions. On the polls of 15 October 2026 no party reaches 76 in 95% of the runs. The ranges are wide for small errors in votes, because with three parties between 21% and 28% the order of exclusion in dozens of divisions turns on a point or two. This paper sets out the design, what comes out, and what the design leaves out.

From a best estimate to a chance

The projection is a single set of national shares, the poll average (Forecasting Australia's House from imperfect polls: an overview), and a single House worked out from it by the division model (The division model and its tests). A reader looking at that House also wants to know how sure we are: whether One Nation could reach 76 if the polls are two points out, how likely a hung parliament is, and which divisions could go either way.

To answer, we need a model of how wrong the inputs might be, pushed through the division model many times. The code is AustraliaForecast.Simulate. Each run is a complete House: every division projected, counted and won.

One run, step by step

A run has five steps.

1. Move each group's national share. For each of the five poll groups (Labor, the Coalition, One Nation, the Greens and others), add a random normal error with a standard deviation equal to the group's root-mean-square final-poll miss, or one point if that is larger. Floor each share at 0.1 and rescale the five to 100.
2. Move each state. In each of the eight states and territories, add to every group's state adjustment an independent random normal error with a standard deviation of 1.5 points.
3. Work out the 2PP. Convert the drawn national primaries to Labor's national 2PP with the 2025 flows, and each state's extra 2PP swing from its drawn adjustments (From primaries to seats: preference flows and the 2PP).
4. Run every division. Project every candidate with equation (2) of The division model and its tests, add a random candidate error, and replay the count. If the final two are Labor and the Coalition, decide the seat by the division's 2PP plus a random division error of 2.5 points.
5. Record. Each grouping's seats, whether the largest party has 76, every combination's total, and every division's winner.

The random number generator has a fixed seed, so the same polls always give the same chances. A change in a published chance always reflects a change in the polls, never a different draw.

The errors

National

Let \(S_g\) be today's national share of group \(g\) and \(e_{g,t}\) the error of the average final poll at past election \(t\) (How wrong have Australian polls been? Primaries, the 2PP and the λ test). The spread is

over the elections from 2007 the model uses, and 1.5 points for a group with no record. In run \(r\),

Table: Standard deviation of the simulated national error, points, from the final polls of 2007–2025.

The error is centred on zero, because the test in How wrong have Australian polls been? Primaries, the 2PP and the λ test found no direction in past misses that would have helped predict the next one. The simulation borrows only their size.

State

Each state's adjustment for each group, \(a_{s,g}\) (equation (3) of The division model and its tests, zero where there are too few state polls), becomes

The state error stands in for the fact that national misses do not fall evenly. It is the same for every division in the state, so it moves whole states together.

Candidate

Each candidate's projected share \(x\) (before rescaling) gets its own error:

A candidate on 5% has a standard deviation of 1.4 points, one on 25% of 3.0 points, one on 40% of 4.2. The candidate error covers what the division model cannot see: local campaigns, sitting members, tactical voting and the spread of real swings around the model's. Its size is a setting, not an estimate.

The 2PP in classic seats

Where the count ends with Labor and the Coalition, the winner comes from

with Labor winning if \(T^{(r)}_d \ge 50\). The 2.5 points stands for the spread of division 2PP swings about the national swing; like the candidate error, it is a setting.

What is not drawn

The preference weights and the 2PP flow rates are held at their 2025 values in every run. The simulation does not ask what happens if One Nation's voters give fewer preferences to the Coalition, or Coalition voters fewer to One Nation. Those uncertainties are real and large (From primaries to seats: preference flows and the 2PP); they reach the simulation only through the candidate error, which moves first preferences and not flows.

What comes out

After 2,000 runs every output is a count.

80% range

Each grouping's seats across the runs, sorted, read at the 10th and 90th percentiles.

Majority and hung parliament

A run gives a majority if its largest party has 76 seats or more; otherwise it is a hung parliament.

Largest party

The share of runs in which a party has more seats than any other. On the every-outcome page a tie for first is shared between the tied parties.

Combinations

The share of runs in which a combination's seats add to 76 or more. The site shows five and lets readers build any other from the stored runs.

Division chances

For each division, the share of runs each grouping wins it.

With 2,000 runs, a reported chance has a Monte Carlo standard error of at most \(\sqrt{0.25/2000} \approx 1.1\) points. Differences of a point or two between chances mean nothing.

Today's figures

On the polls to 11 October 2026:

Table: The 2,000 runs, 15 October 2026. "Largest party" shares ties.

No party has 76 in 95% of the runs. Three features of the table are worth reading closely.

The ranges are wide. One Nation's 80% range spans 21 seats and Labor's 22, for national errors of 1.25 and 1.85 points. In a House where 88 divisions are projected to end as Labor against One Nation, a point moved between the two moves many seats at once.

The runs lean away from One Nation. One Nation's mean is 59.8 seats, four below the projection, and the Coalition's 15.2, two above. The projection gives One Nation many divisions it wins narrowly: of the 19 divisions where the projected winner's chance is under 60%, eight are One Nation's, and 22 of its 64 are under 75%. Random errors take more of those away than they add elsewhere.

Labor and One Nation trade seats. Across the runs the correlation between Labor's and One Nation's seat counts is −0.74. Because the five shares are rescaled to 100, a run that raises one group lowers the others, and in this House the seats one gains are mostly the other's.

The chance of a majority for a combination:

Table: Combinations, from the 2,000 runs. No agreement exists for any of them; they are shown for the arithmetic.

The first two are close to even. The 80% range of the Coalition and One Nation together is 63 to 87 seats, and of Labor, the Greens, the independents and Centre Alliance together 63 to 86. Which side could form a government turns on the same few dozen three-way contests.

In the divisions, the projected winner's chance is 95% or more in 57, 75–95% in 45, 60–75% in 29, and under 60% in 19.

Why the errors are drawn this way

One draw per group, then a rescale. The British simulation replays whole past misses, so that parties' errors move together as they have before (Simulating polling error: where the chances come from). The Australian one draws each group's error independently and rescales. Two reasons. First, the Australian record has seven elections for the two largest parties and three for One Nation and others, and the pattern of misses changes: Labor and the Coalition missed in opposite directions at five of the seven, but in the same direction in 2007 and 2016. Second, a pattern-based draw can only produce errors in the directions past patterns span, and no past pattern contains a One Nation at 26%. Independent draws give every group its own uncertainty. The rescale then adds the one constraint every run must respect, that shares add to 100, and with it a negative correlation between groups.

A state level. A national miss need not fall evenly across states, and the 2022 → 2025 test showed that state swings help when they are right (The division model and its tests). The state error moves all divisions in a state together, which widens the ranges more than independent division errors of the same size would.

A candidate error that grows with share. A fixed error would swamp small candidates and understate the uncertainty for large ones. The form \(1 + 0.08x\) gives a minor candidate about a point and a leading candidate three to four.

Before polling day

The election-day forecast runs the same simulation, with 1,000 runs, on shares pulled part of the way back towards 2025, and adds drift for the days left. Each group's national standard deviation becomes

where \(\delta(d)\) is the drift for \(d\) days to go (Polling day: how far Australian polls move). At 583 days \(\delta\) is 2.7 points for a group at 20%, and less for small groups: the factor \(\sqrt{S_g/20}\) gives a group at 5% half the drift of one at 20%. On those settings Labor wins a majority in 84% of the polling-day runs, and the chance of a hung parliament falls to 15%. That forecast rests on the pull-back, which for One Nation is an assumption.

What the simulation does not capture

Preference uncertainty. The weights and flows are fixed at 2025. A change in how Coalition voters rank One Nation against Labor would move dozens of seats together, and the simulation has no term for it.

Correlation from history. The national errors are independent before the rescale. At real Australian elections Labor and the Coalition have usually missed in opposite directions; the rescale produces some of that, but not its size.

Local error that is not random. Candidate errors are independent from division to division. Real errors of a swing model run in the same direction across similar places: outer suburbs, regional Queensland, inner cities.

New candidates and sitting members. The field is the 2025 field. A strong new independent or a member who has changed party is outside the model.

Everything outside polling error. Events, leadership changes and the date of the election.

Calibration

A probability forecast is well calibrated if events given 20% happen about 20% of the time. A handful of elections cannot test that. What can be said is that the design reproduces, by construction, the size of the national misses since 2007, and that its division layer was tested from 2022 to 2025 and on six earlier election pairs (The division model and its tests). It does not reproduce misses larger than any in the record, and it does not vary preferences. We think the national ranges are about right for the size of polling error Australia has seen, and that the combination chances are the least secure numbers, because they depend on the shape of the error across many divisions at once.

References

Australian Electoral Commission, House of Representatives results and distributions of preferences, 2004–2025 (CC BY 4.0).

Jennings, W. and Wlezien, C. (2018). Election polling errors across time and space. Nature Human Behaviour, 2(4), 276–283.

Shirani-Mehr, H., Rothschild, D., Goel, S. and Gelman, A. (2018). Disentangling bias and variance in election polls. Journal of the American Statistical Association, 113(522), 607–614.

Figures are as of 15 October 2026. The site's pages update with each new poll.