Australia Next federal electionby 20 May 2028in 546 days Hung parliament 96% Labor majority 1% Calendar Track record Research
Research · Australia

From primaries to seats: preference flows and the 2PP

Abstract. The forecast works out Labor's 2PP from the primaries with the share of each group's votes that reached Labor in the AEC's 2025 count, which came within about a point of the real 2PP at every election from 2007 when tested with the previous election's flows. Each division's count is replayed with preference weights between seven groups, fitted by a Luce choice model on all 826 exclusions of 2025; on the actual 2025 votes it picks 149 of 150 winners. The weakest number is how Coalition voters rank One Nation against Labor.

Polling Forecast · updated 15 October 2026 · 9 min read · Australia, preferences, two-party-preferred, distribution of preferences, Luce model, minorisation-maximisation, preference flows, One Nation, Greens, Dickson

Summary

Australian elections are decided after preferences. This paper describes the two places where the forecast needs to know how preferences flow. The first is the national two-party-preferred (2PP): Labor's share against the Coalition after preferences, which the forecast works out from the primary votes with the share of each group's votes that reached Labor in the AEC's 2025 count. On 2025's own primaries this gives the AEC's 55.2 exactly; on each past election's primaries with the previous election's flows it came within about a point every time from 2007 to 2025; on today's polls it gives Labor 52.2. The second is the count in each division, which is replayed with preference weights between seven groups of candidates, fitted by a Luce choice model on all 826 exclusions in the AEC's 2025 distributions of preferences. Run on the actual 2025 first preferences, the replayed count picks 149 of the 150 winners; with the 2022 weights instead, out of sample, 146.

The weakest number in the table is the one that matters most this term: how Coalition voters rank One Nation against Labor when the Coalition candidate is out. In 2025 few divisions put that to the test.

Why preferences matter more than they did

Voters number every candidate. If no candidate has more than half the formal first preferences, the candidate with the fewest is excluded and their ballots pass to the next preference shown, and so on until one candidate has a majority. The AEC counts every division through to the last two, giving the two-candidate-preferred (2CP) result, and separately distributes every division as Labor against the Coalition, giving the 2PP even where the real final two were different.

For most of the past fifty years the 2PP summed up an election, because most divisions ended as Labor against the Coalition and preferences mattered only in deciding which of the two won. In October 2026 the polls put One Nation level with Labor on first preferences. In a division where three candidates are within a few points of each other, who is excluded third decides the seat, and the national 2PP says nothing about who that is. The forecast therefore uses the 2PP only where the final two are Labor and the Coalition, and replays the count everywhere else.

The 2PP from primaries

The formula

The AEC publishes, for every party, how its voters' preferences split between Labor and the Coalition in the 2PP count. Summed by group and divided by each group's national first preferences, these give the share \(r_g\) of group \(g\)'s votes that reached Labor. Labor's national 2PP for primaries \(S_p\) is then

\[ T = \frac{S_{\text{ALP}} + \sum_{g \neq \text{ALP}} r_g\, S_g}{\sum_p S_p} \times 100. \tag{1} \]

The rates from the 2025 count are:

Group Share of first preferences reaching Labor, 2025
Greens 88.2%
Others (independents and minor parties) 54.5%
One Nation 25.5%
Coalition 0.2%

Table: The 2025 2PP flow rates by group, from the AEC's 2PP flow file.

The Coalition's own rate is not zero because, in three-cornered contests, some Liberal or National votes reach Labor before they reach the other Coalition candidate. On the 2025 primaries equation (1) gives 55.22, the AEC's figure, by construction.

A worked example: today

On the average of 15 October 2026 (Labor 27.67, One Nation 26.50, the Coalition 21.42, the Greens 13.17, others 11.25):

\[ T = 27.67 + 0.002 \times 21.42 + 0.255 \times 26.50 + 0.882 \times 13.17 + 0.545 \times 11.25 \approx 27.67 + 0.05 + 6.76 + 11.61 + 6.14 = 52.2. \]

Labor's 2PP has fallen three points since 2025, although its primary has fallen by seven, because the Greens and others, whose voters prefer Labor, have held their share, and because most of Labor's lost primary vote has gone to One Nation, a quarter of whose votes still reach Labor.

The test

The rates move from one election to the next, and the forecast can only use the last election's. To see how much that costs, each election's actual primaries were converted with the previous election's rates and compared with the AEC's 2PP:

Election From primaries, previous election's flows AEC 2PP Difference
2007 52.7 52.7 0.0
2010 50.4 50.1 +0.3
2013 45.5 46.5 −1.0
2016 49.4 49.6 −0.2
2019 49.3 48.5 +0.8
2022 51.2 52.1 −0.9
2025 54.8 55.2 −0.4

Table: The 2PP conversion tested out of sample, Labor's 2PP, points.

The conversion was within about a point every time. These gaps are the part of a 2PP miss that comes from flows changing, not from the polls. They are smaller than the pollsters' own 2PP misses in 2010, 2019 and 2025 (How wrong have Australian polls been? Primaries, the 2PP and the λ test).

How stable the flows are

Election Greens → Labor One Nation → Labor Others → Labor
2004 81% 44% 44%
2007 80% 47% 45%
2010 79% 45% 42%
2013 83% 45% 47%
2016 82% 50% 49%
2019 82% 35% 46%
2022 86% 36% 50%
2025 88% 26% 55%

Table: Share of each group's first preferences that reached Labor in the AEC's 2PP count.

The Greens' flow to Labor has been steady and rising. Others' has risen from the low forties to 55%. One Nation's has fallen from 44 to 50% (2004–2016) to about 35% (2019, 2022) and 26% (2025). That fall matters now: One Nation's voters are a quarter of the electorate, and a flow 10 points stronger or weaker would move Labor's 2PP by about 2.6 points. We use the last election's flows, which is also what most pollsters who use "previous-election" preferences do.

The pollsters' own 2PP

The forecast does not use the pollsters' published 2PP, but the site shows it beside ours. Pollsters work it out in one of two ways: with the flows of the last election, as we do, or by asking respondents for their preference ("respondent-allocated"). Roy Morgan publishes both; Newspoll uses previous-election flows; Essential publishes a "2PP+" that keeps an undecided share. In the 21-day window on 15 October 2026:

Pollster Date Labor v Coalition Labor v One Nation Ours from its primaries
Roy Morgan 11 Oct 53.0 (53.5 on 2025 flows) 54.5 52.8
YouGov 6 Oct 52.0 53.0
RedBridge 2 Oct 46.0 52.0
Essential 28 Sep 48–46 with undecided

Table: Published 2PP figures in the current window. Resolve and Newspoll published no 2PP with these polls.

The pollsters' average Labor v Coalition figure, leaving out Essential's because it keeps undecided voters, is 50.3, against our 52.2. Most of the difference is RedBridge's 46, which implies a much stronger flow from One Nation to the Coalition than in 2025. Morgan's own figure on 2025 flows (53.5) is 0.7 points above ours from the same primaries.

The weights between groups

What the count needs

To replay a count we need to know where an excluded candidate's votes go when several candidates remain. The AEC's distribution of preferences records every exclusion in every division: who was excluded, and how many votes went to each candidate still in the count. That is the evidence.

Candidates are put into seven groups: Labor, the Coalition, the Greens, One Nation, independents (Centre Alliance and Community Strong Australia included), right-leaning minor parties, and other minor parties. A party counts as right-leaning if its voters gave the Coalition more preferences than Labor (under 42% to Labor in an AEC 2PP flow file); Katter's Australian Party, the United Australia Party, Trumpet of Patriots and Family First are among them. The other minor parties are the rest.

The Luce model

When a candidate of group \(a\) is excluded, each candidate \(i\) still in the count receives a share of the votes

\[ \pi_i = \frac{w(a, g_i) / n_{g_i}}{\sum_{k \in R} w(a, g_k) / n_{g_k}}, \tag{2} \]

where \(R\) is the set of candidates remaining, \(g_i\) is candidate \(i\)'s group, \(n_g\) is the number of remaining candidates of group \(g\) (so two independents share the independents' weight), and \(w(a, b)\) is the weight from group \(a\) to group \(b\). This is a Luce choice model (Luce, 1959): the odds between any two receiving groups do not depend on who else is left.

The weights are fitted by minorisation-maximisation (Hunter, 2004). For each source group \(a\), with \(E_a\) its exclusions, \(v_{e,b}\) the votes group \(b\) received at exclusion \(e\), and \(V_e\) the total transferred, each iteration sets

\[ w(a, b) \leftarrow \frac{\sum_{e \in E_a} v_{e,b}}{\sum_{e \in E_a} V_e \big/ \sum_{c \in e} w(a, c)}, \tag{3} \]

with a floor of \(10^{-4}\), then divides every weight from \(a\) by the largest weight from \(a\) to another group, so that the largest is 1. The fit runs 200 iterations. The weight from a group to itself (an independent's votes going to another independent) is fitted the same way but left out of the scaling, so it can exceed 1.

The 2025 weights

Fitted on all 826 exclusions in the 2025 distributions of preferences:

From \ to Labor Coalition Greens One Nation Independents Right minor Other minor
Labor 0.61 0.31 0.96 0.21 1.00 0.72 0.61
Coalition 0.30 1.21 0.11 1.00 0.68 0.08 0.12
Greens 0.93 0.21 0.56 0.19 1.00 0.25 0.52
One Nation 0.19 0.60 0.18 0.37 0.49 1.00 0.21
Independents 0.86 0.56 0.80 0.68 1.71 0.69 1.00
Right minor 0.24 0.40 0.18 1.00 0.43 0.44 0.44
Other minor 0.84 0.31 0.73 0.86 1.00 0.74 1.12

Table: Preference weights, 2025. The diagonal is the weight between two candidates of the same group.

Reading a pair off the table is a matter of equation (2) with two groups left. Greens votes split between Labor and the Coalition \(0.93 / (0.93 + 0.21) = 82\%\) to Labor. One Nation votes split between the Coalition and Labor \(0.60 / (0.60 + 0.19) = 76\%\) to the Coalition. Coalition votes split between One Nation and Labor \(1.00 / (1.00 + 0.30) = 77\%\) to One Nation.

The table shows the familiar pattern. Greens preferences go mostly to Labor and independents; One Nation's to right-leaning minor parties and the Coalition; the Coalition's to One Nation and independents well ahead of Labor and the Greens; independents' to other independents first, then Labor and the Greens.

A worked example: Dickson

In Dickson, on today's projection, the count reaches a point where the Coalition (28.6% after earlier transfers), Labor (40.5%) and One Nation (30.9%) remain. The Coalition is excluded. By equation (2), with weights 1.00 to One Nation and 0.30 to Labor, One Nation receives \(28.6 \times 1.00/1.30 = 22.0\) points and Labor \(28.6 \times 0.30/1.30 = 6.6\). One Nation finishes on 53.0 and Labor on 47.0 (the figures on the division page, from unrounded weights and shares). Labor would win if fewer than about two-thirds of those Coalition votes went to One Nation: at 60:40, for example, Labor would finish on 51.9.

Stability of the weights

The same model fitted separately on each election's exclusions gives, for the pairs that decide most seats:

Election Exclusions Greens → Labor v Coal. One Nation → Coal. v Labor Coal. → One Nation v Labor Ind. → Labor v Coal. Other → Labor v Coal. Right → Coal. v Labor
2004 791 78% 58% 42% 43% 56% 71%
2007 754 76% 59% 77% 50% 55% 66%
2010 549 76% 59% 82% 46% 49% 69%
2013 888 79% 54% 70% 50% 56% 56%
2016 694 79% 50% 74% 53% 54% 67%
2019 754 77% 66% 43% 50% 74% 69%
2022 901 80% 66% 64% 57% 77% 66%
2025 826 82% 76% 77% 61% 73% 62%

Table: Share going to the first-named group when only the two named groups remain, from weights fitted on each election alone.

Greens preferences have been steady, between 76% and 82% to Labor. One Nation's have moved towards the Coalition since 2016, from 50% to 76%. The Coalition's flow to One Nation against Labor has swung between 42% and 82%. One Nation stood in few divisions from 2007 to 2016, so that column rests on few exclusions for those years, and the jumps say as much about thin evidence as about voters.

Pooling every election's exclusions instead of using the last election's alone was tested on the election-to-election replays (The division model and its tests): it got 3 more winners right over five election pairs. We keep the last election's weights, because pooling would pull One Nation's flows back towards their 2004–2016 levels, which the 2PP flow table shows to be out of date.

The count tested on 2025

Run on the actual 2025 first preferences in every division:

Weights Winners right Final pair right Average error on the winner's 2CP
2025 (in sample) 149 of 150 147 1.35 points
2022 (out of sample) 146 of 150 143 1.47 points

Table: The replayed count on the 2025 first preferences. The 2CP error is measured where the final pair is right.

The in-sample result shows that a group-level count, with no candidate's own how-to-vote card and no individual ballot paths, reproduces almost every 2025 outcome. The out-of-sample result shows what is lost when the weights are a cycle old: three more winners wrong and four more wrong pairs. The seat totals from the out-of-sample count were Labor 93, the Coalition 40, independents 13 and the Greens 2, against the real 94, 43, 10 and 1.

Where the preference model is weakest

Coalition voters choosing between One Nation and Labor. In 2025 few divisions reached a point where the Coalition was out and both One Nation and Labor remained. The weight that decides the 88 projected Labor against One Nation contests and the 19 Coalition against One Nation contests rests on that thin evidence. The pollsters' own Labor v One Nation 2PP figures (52 to 54.5 for Labor in the current window) suggest a weaker flow from the Coalition to One Nation than ours.

Flows when a party quadruples. The 2025 weights describe One Nation voters when One Nation had 6.4% of the vote. Voters who have moved to it since 2025, many of them former Coalition voters, may give their preferences differently.

Groups, not candidates. Every independent is treated alike, though a "teal" independent's voters and a rural independent's voters give preferences very differently. Within-group splits follow equation (2) and give each candidate of a group the same weight.

What we don't know yet: limitations of the Australian forecast returns to these points.

References

Australian Electoral Commission, distributions of preferences and 2PP preference flows by party, 2004–2025 (CC BY 4.0).

Hunter, D. R. (2004). MM algorithms for generalized Bradley–Terry models. The Annals of Statistics, 32(1), 384–406.

Luce, R. D. (1959). Individual Choice Behavior: A Theoretical Analysis. New York: Wiley.

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

How to cite

Polling Forecast (2026). From primaries to seats: preference flows and the 2PP. Retrieved 21 November 2026 from https://www.pollingforecast.com/research/australia-preferences. Live figures change as new polls arrive; the date above is when the text was last revised.

You may also like