Previous

Landslide

20

The polls: Underestimating Labor, overestimating the Coalition

Murray Goot

Abstract

All 10 national opinion polls exaggerated the Coalition’s position and underestimated Labor’s. Had they predicted a Coalition victory, even a hung parliament, their performance would have generated more consternation. From one poll to another, there was little variation in the estimates for either Labor or the Coalition. For this, herding is one possible explanation. But there are others: survey errors, which are common to most of the polls; the treatment of the ‘undecided’; and shared assumptions that, in a cost-of-living ‘crisis’, the Coalition’s vote was likely to rise and that, if Labor won, it was unlikely to win by much. The attempt to predict the vote seat by seat using multilevel regression with post-stratification (MRP) was equally unimpressive. While Labor gained 11 seats from the Coalition, the MRP predicted that it would win only seven. The continuing success of the electoral pendulum in predicting seats shares, the rise of Independents notwithstanding, remains noteworthy.

Keywords: opinion polls; Australian elections; electoral pendulums; MRP; herding

In the media’s coverage of campaigns, polling is central. The calling of the horserace generates considerable interest (not least among political journalists and politicians), controversy (different polls calling it differently) and widespread publicity—for the media that publishes them and for the pollsters who produce them. If getting the figures ‘right’ brings kudos, getting them ‘wrong’ causes handwringing, finger-pointing and reputational damage.

The 2025 campaign involved more pollsters than ever. The internet has made polling cheaper; online polls cost as little as one-third that of a telephone (computer-assisted telephone interviewing) poll, which was the usual polling mode until 2016. There were no formal barriers to conducting polls, no audits of any kind or disclosure rules; most pollsters were not members of the Australian Polling Council (2025), the only body that sets minimum disclosure standards. And publishing polls had never been easier: in the absence of a media contract, pollsters could always publish on the internet. Since pollsters can sponsor more than one poll (two did this time) and since they can invoke different ways of judging their performance—predicting which side will win or will come closest to winning (a hung parliament), predicting the 2PP vote or the parties’ first-preference shares, and so on—participating may be worth the risk.

During the final days of the 2025 campaign no fewer than 10 pollsters published their predictions of the national vote. Some were published (though not always commissioned) by the print media: Newspoll in the Australian; Freshwater in the Australian Financial Review; Resolve in the Age and The Sydney Morning Herald; and RedBridge/Accent in News Corp’s metropolitan mastheads. Some were published online: Essential in the Guardian and Ipsos in the Daily Mail—both British-owned mastheads. And some were published by pollsters with no assured media outlet: DemosAU, Roy Morgan, Spectre and YouGov. There were attempts also to measure the vote in individual seats (YouGov, especially its MRP; but also Compass, DemosAU, Freshwater, JWS, KJC, the New England Times and UComms) or in clusters of seats (DemosAU, RedBridge/Accent)—all except RedBridge/Accent’s financed by the pollsters themselves.1

How well did the polls perform, collectively as well as individually, in estimating the distribution of first preferences and the 2PP vote? How good was the polling based on MRP? And which approach offered the better guide to the outcome: the approach that assumed that if Labor increased its national vote share it would be returned, the bigger any increase the more comfortable its majority; or the approach that assumed that with the rise of the Independents, in particular, the election was best understood as 150 separate contests, making the MRP one’s best bet?

I start by sizing up the errors made in estimating the 2PP vote: most, it turns out, stemmed from the polls’ failure to estimate each party’s first preferences. In trying to explain this, I consider the importance of sampling and the handling of the ‘undecided’ vote, as well as how the polls allocated preferences. Different ways of measuring the errors produced different leader boards. In single seats, measures of success also vary, though some polls were poor by any measure. The third task, seeing whether the national polls’ estimates of the two sides’ seat shares outperformed the estimates offered by the MRP, proved the most straightforward task of all.

The national polls

How good were the national polls? Since elections are one of the few public tests they face, and they have much at stake in the outcome, pollsters are keen to set the terms of these tests, to indicate how they should be applied and to make sure their own judgements about best and worst prevail. This means getting in early; first impressions can be lasting ones.

Weeks before the official count was finished, several pollsters had issued their own assessments of how they had done, and sometimes of how their competitors had done. Others did so in response to journalists’ questions. The focus was on how well they had called the 2PP vote. Doing this better than anyone else meant winning ‘line honours’, as The Poll Bludger’s William Bowe (2025a) put it. The question of how well the polls had estimated first-preference votes was a secondary matter.

The two-party-preferred vote

Announcing, three days after the election, that it had ‘aced’ the election, Roy Morgan made it clear that its polling would come out well no matter the final vote. That the Albanese government would be ‘re-elected with an increased majority’ was an outcome Roy Morgan had ‘predicted last week and for the previous seven weeks since mid-March when the ALP took a decisive lead’. Its final poll ‘showed the ALP ahead 53% to 47% based on respondent allocated preferences’ but ‘ahead 54% to 46%’—‘in line with the projected final result’, it emphasised—‘based on how [preferences] flowed at the previous Federal Election’. With 63 per cent of the vote counted, Labor was ahead 54.6 to 45.4, and the result was expected to be ‘close to’ 54–46 (Roy Morgan 2025b); the actual result would be different: 55.2–44.8 per cent.

Roy Morgan had not tipped a 54–46 result. Guided by how respondents said they would allocate preferences, it had tipped a 53–47 result. Readers had to read to the end of Roy Morgan’s press release to learn that ‘based on how Australians voted at the 2022 Federal Election’, the 2PP result favoured Labor, 54–46 (Roy Morgan 2025a). Not only did Roy Morgan’s final report privilege respondent allocation over the 2022 allocation; so, too, over the previous seven weeks did every one of its reports. Roy Morgan’s decision to publish two sets of figures bought it cover: it allowed it to cite whichever set of figures came closer to the real result. Publishing two or more sets of figures, based on different preference allocations (Goot 2021: Appendix) or different survey modes (Mansillo and Jackman 2020: 138), was a practice Roy Morgan had used in the past.

Resolve’s Jim Reed was another pollster keen to jump in. With 77 per cent of the vote counted, and the 2PP vote favouring Labor 54.1–45.9, Reed declared that Resolve and RedBridge had produced the two most accurate estimates of the 2PP vote; overlooked were the Roy Morgan poll (described as ‘unpublished’) and the Spectre poll, both of which had produced the same estimates. The ‘two least accurate’ were Freshwater (51.5–48.5) and Ipsos (51–49).

According to Reed (2025), it was ‘those polls taken a week out that [had] proven most accurate’. Singled out were ‘the Resolve Political Monitor and RedBridge, both of whom called the TPP result within their error margins’. Since the five polls conducted three-to-six days out from the election produced an average 2PP vote of 52.6–47.4, and the five polls conducted one or two days out produced an average 2PP vote of 52–48, there was not much in it; with ‘error margins’ included, there was nothing in it. ‘Pollsters left red-faced by Labor’s powerful late surge’ was the very different headline over The Sydney Morning Herald’s post-election review (Knott 2025). By the end of the count (55.2–44.8), all the polls had proved to be out by much more than Reed had imagined (Table 20.1).

Table 20.1 Accuracy of the national polling, House of Representatives, 2025 (per cent)

Error

(primary)

Error

2PP (ALP)

Poll

Fieldwork

Days to election

Sample size

Mode

ALP

Coalitiona

Greens

PHON

TOP

Other

DKb

ALP–Coalition

Avg. (5)d

RMSE (5)d

Est.

Error

Roy Morgan

28 April – 2 May

1

1,368

I+CATI+FtF

33.0

34.5

13.5

6.5

2.0

10.5

[5]

–4.3

1.6

3.56

53

–2.2

Freshwater

29 April – 1 May

2

2,055

I

33.0

37.0

12.0

8.0

n.a.

10.0

n.a.

–6.8

2.8

11.60

51.5

–3.7

Ipsos

28 April – 1 May

2

2,574

I+CATI

29.5

34.7

12.6

8.4

2.1

12.6

[5]

–8.0

2.1

7.73

51

–4.2

Newspoll (Pyxis)

28 April – 1 May

2

1,270

I

33.0

34.0

13.0

8.0

n.a.

12.0

[4]

–3.8

1.8

3.92

52.5

–2.7

YouGov

24 April – 1 May

2

3,003

I

31.1

31.4

14.6

8.5

2.5

11.9

n.a.o.

–3.1

1.8

4.40

52.2

–3.0

DemosAU

27–30 April

3

4,100

I

31.0

33.0

12.0

9.0

n.a.

15.0

[7]

–4.8

1.5

4.24

52

–3.2

RedBridge/Accent

24–29 April

4

1,011

I

34.0

34.0

12.0

8.0

n.a.

12.0

[5]

–2.8

1.5

3.36

53

–2.2

Spectre

24–28 April

5

2,000

I

31.0

34.0

14.5

9.5

n.a.

11.0

[9]

–5.8

3.0

9.74

53

–2.2

Resolve

23–28 April

5

2,010

I+CATI

31.0

35.0

14.0

7.0

n.a.

13.0

n.a.o.

–6.8

2.2

6.16

53

–2.2

Essentialc

24–27 April

6

2,163

I

32.0

34.0

13.0

10.0

2

9.0

[5]

–4.8

2.6

8.24

52.1

–3.1

Average

–

–

–

–

31.9

34.2

13.1

8.3

–

12.6e

–

–

–

–

52.3

–

Election

3 May

–

–

–

34.6

31.8

12.2

6.4

1.9

13.1

–

–

–

–

55.2

–

Error

–

–

–

–

–2.7

+2.4

+0.9

+1.9

–

–2.4e

–

–

–

–

–2.9

–

Notes: n.a. = not available; n.a.o. = not an option; RMSE = root mean squared error; I = internet panels (Roy Morgan, 33 per cent; Ipsos, 67 per cent; Resolve, n.s.); CATI = computer-assisted telephone interviewing (Roy Morgan, 33 per cent; Ipsos, 33 per cent; Resolve, n.s.); FtF = face-to-face (Roy Morgan, 33 per cent). a Liberal Party plus Liberal National Party (Queensland) plus National Party plus Country Liberal Party (Northern Territory). b Respondents `undecided’ about their first preference after the `leaner’, subsequently deleted or distributed. c Undecided (4.8 per cent) on the two-party vote ignored for purposes of comparability; Essential’s two-party figures were Labor, 49.6 per cent; Coalition, 45.6 per cent; and undecided, 4.8 per cent. d Based on five blocks of respondents: ALP, Coalition, Greens, PHON, TOP and Other. e Includes TOP; Figures in bold text are the best scores on this measure; those in italics are the worst scores on this measure.

Sources: AFR (2025); Beaumont (2025a); Butler (2025); Crowe (2025); DemosAU (2025); Ipsos Australia (2025); Pyxis Polling and Insights (2025); Roy Morgan (2025a); Spectre Strategy (2025); Wikimedia Foundation (2025); YouGov (2025b).

Immediately after the vote, Michael Turner, CEO of Freshwater, conceded that he had underestimated Labor’s 2PP vote, but insisted that there had been ‘a late swing among “soft” or undecided voters’ that his poll and others found ‘hard to pick up’ (Turner 2025). But the evidence he adduced suggests otherwise. Of the polls he thought best—RedBridge, followed by Roy Morgan and Resolve—RedBridge had completed its fieldwork two days earlier than Freshwater, Morgan had completed its fieldwork a day later and Resolve, a day earlier (Table 20.1).

Writing for the masthead that had commissioned the Freshwater poll, journalist Michael Read recycled Turner’s views. As for the almost equally wayward Ipsos poll, Read relied on the words of the company’s research director: after being ‘a little too progressive in the two-party preferred’ in ‘a couple’ of earlier elections, Ipsos had ‘overcorrected’ (Read 2025).

Preferences

How the polls allocated preferences was widely seen as the key to their errors. In allocating preferences, different pollsters adopted different methods. RedBridge/Accent used the preference flows from 2022. Newspoll and Essential used flows from ‘previous elections’, including State elections. Ipsos, Roy Morgan, Resolve and Essential made two separate calculations—one based on preference flows in 2022; the other based on the respondents’ own allocations—though they did not always publish the two. YouGov and Spectre ‘blended’ the two. RedBridge/Accent did as well, but only in relation to the Trumpet of Patriots (TOP); otherwise, it says, it based its allocation on the 2022 flows.

Did any of this make any difference? Roy Morgan’s figures did vary, depending on whether it used respondent allocation or the 2022 flows. Ipsos and Essential, however, reported no difference. When the Conversation’s psephologist Adrian Beaumont plugged the 2022 preference flows into 2025’s first preferences, he came up with ‘a Labor two-party lead of 55.3–44.7, [virtually] the same as the actual result’ (Beaumont 2025b).

The extent to which the pollsters did what they said they did is difficult to determine. Subtracting Labor’s 2PP vote from its first-preference vote and dividing this number by the proportion of respondents who voted for someone other than Labor or the Coalition show that every pollster’s allocation favoured Labor; since the Greens’ vote accounted for more than one-third of all the preferences, this is not surprising. On YouGov’s figures, 56 per cent of all preferences went to Labor; on Resolve’s figures, it was 65 per cent. But given that YouGov was distributing the preferences of 37.5 per cent of its sample (those who were not giving their first preference to Labor or the Coalition), and Resolve was distributing the preferences of 34 per cent, the difference between the two in 2PP terms amounted to just 1 percentage point. These were the two outliers. Among the others, the range was narrower: 58 per cent (DemosAU) to 63 per cent (Spectre). On the day, Labor won just over 61 per cent of the second preferences (derived from Table 20.1), the midpoint of the polls’ estimates.

Regardless of when they were conducted or how they allocated preferences, all the polls underestimated Labor’s share of the 2PP vote (Table 20.1). On average, the error was not ‘almost’ as great as it was in 2019 (Beaumont 2025b), when all the polls picked the wrong winner; the average was just as great. Two of the polls’ estimates would have seen Labor’s 2PP vote slip back to 51 per cent (Ipsos) or 51.5 per cent (Freshwater) from 52.1 per cent, its 2022 level; one would have seen it stay the same (Essential); the others would have seen it increase to 52 per cent (DemosAU), 52.2 per cent (YouGov), 52.5 per cent (Newspoll) or 53 per cent (Roy Morgan, RedBridge/Accent, Spectre, Resolve).

Underestimates of Labor’s 2PP share (55.2 per cent) ranged from 4.2 to 2.2 percentage points—from estimates that might have seen Labor lose its majority to estimates that might have seen its majority increase. By this measure, line honours were shared. Only by going to a second decimal point—mercifully, avoided—would the stewards have been able to separate Roy Morgan from RedBridge/Accent, Resolve or Spectre. But even a best of 2.2 was not particularly good (Goot 2021: Appendix).

One way of rescuing the reputation of the polls, or at least some of them, Beaumont (2025b) suggests, is to look through their back catalogue for figures that were closer than their final figures. This is a suggestion best ignored. Polls are time sensitive: ‘right at the time’, as pollsters like to say, not necessarily right sometime later. Beaumont’s suggestion would open the way to cherry-picking. What if another pollster had produced a poll, say, four months earlier that was close to the result? Would their earlier poll have made them one of the best, regardless of their subsequent polls? His suggestion would also mean that changes in the polls could not be used as evidence of a shift in the final week.

Before 2PP became the dominant metric, honours may have been decided by which poll best estimated the gap, on first preferences, between the government and the Opposition (Beed 1977: 222). On this measure, the winner would have been RedBridge/Accent: it had the two level-pegging, in a photo finish with YouGov (the Coalition 0.3 of a point ahead); Roy Morgan (which had the Coalition 1.5 points ahead) would have come fourth, Spectre (three points) ninth and Resolve (four points) last (Table 20.1).

For the pollsters, the 2PP vote has been a boon (Goot 2021: 53). As Table 20.1 shows, the error based on how close each of the polls came to estimating the gap in first preferences between Labor and the Coalition was larger than the error based on the gap in the 2PP. For Roy Morgan, Freshwater and Ipsos, the error on first preferences was more than twice as large; for Spectre and Resolve, it was more than three times as large.

First preferences

The most important reason for Labor’s 2PP vote being underestimated was not the formula used for allocating preferences; it was that Labor’s first-preference vote was underestimated. The Coalition’s vote was overestimated. These errors were compounded by the overestimation of the vote for Pauline Hanson’s One Nation (PHON), only partly offset by overestimates for the Greens.

By the end of counting, Labor had secured 34.6 per cent of the first preferences. The average estimate across the polls was 31.9 per cent—an underestimate of 2.7 points. Not a single poll estimated a Labor vote higher than 34.6 per cent. While the Coalition secured 31.8 per cent of the first preferences, the polls’ average estimate was 34.2 per cent—an overestimate of 2.4 points. Only YouGov (31.4 per cent) came close to estimating the Coalition’s vote. The median error in estimating the gap between Labor and the Coalition’s first-preference vote was 4.8, which is roughly in line with the performance of the polls since the 1940s internationally (Jennings and Wlezien 2018: 280, Fig. 2c).

According to Bowe (2025b), a ‘large component of the pollsters’ failure’ to estimate Labor’s share of the 2PP vote ‘lay in an expectation … that preference flows to Labor would not match those of 2022’. Not so. Since the polls underestimated Labor’s first preferences by 2.7 points, on average, and its share of the 2PP vote by 2.9 points, the error in estimating Labor’s 2PP vote was almost entirely due to the errors in estimating its first-preference vote.

In addition to showing how well the polls estimated the first preferences of the parties of government, Table 20.1 shows how well the polls estimated the vote for the Greens, PHON and ‘Others’—the other parties and Independents on which each of the polls reported. The best-performed polls were those by RedBridge/Accent and DemosAU; their average error, across each of the five, was 1.5 percentage points. Next came Roy Morgan (1.6). The worst performed, with errors up to twice as large, were Spectre (3.0), Freshwater (2.8) and Essential (2.6). Internationally, since the 1940s, average errors have been less than two points (Jennings and Wlezien 2018: 278).

Another measure, less widely used, is the root mean squared error of these errors. This punishes polls for big errors and rewards polls for errors that are only small. Here, honours again go to RedBridge/Accent (with a score of 3.36 percentage points), followed by Roy Morgan (3.56) and Newspoll (3.92). The wooden spooners are Essential (8.24), Spectre (9.74) and Freshwater (11.60).

While two of the four (Roy Morgan and RedBridge/Accent) that did best on the 2PP vote were also among the best on these two measures, Resolve and Spectre were among the best on neither of these measures.

Sampling errors and handling the ‘undecideds’

Since almost all the polls erred in the same ways, what are the likely sources of these errors? Coverage error is one. The sampling frames used by online panels may not include all the relevant population—most obviously, voters who are not online. They recruit potential respondents by invitation, and the numbers recruited are tiny (Pennay et al. 2019: 48). To mitigate these problems, some pollsters added sampling frames: Ipsos and Resolve used telephone interviews; Roy Morgan, telephone and face-to-face interviews. The success of these strategies is unclear (Pennay et al. 2019: 49–50).

Another source of error—the only one that pollsters are happy to flag—is sampling variance (‘sampling error’): the figure (‘±3.2 per cent’, depending on sample size) that assures readers that 95 per cent of the time polls come within a few percentage points of the real figures. These calculations, however, assume probability sampling, in which every voter has an equal (or at least known) chance of being selected. However, from already unrepresentative online panels, pollsters select not at random but by quota: a certain age distribution, equal numbers of men and women, and on so (Pennay et al. 2019: 50).

If sampling variance explained the errors, instead of all the polls underestimating Labor’s 2PP vote, roughly as many polls would have overshot the real figure as undershot it. It was a similar case for estimates of Labor’s first preferences (all too low), the Coalition’s (all but one too high), One Nation’s (all but one too high) and Others’ (all but one too low). While some estimates of the Greens’ vote were correct, and some too high, none was too low. It is this lack of ‘dispersal’ that shows that sampling variance is not the answer.

Non-response is a third potential source of error. Even if the samples drawn from the panel providers (or other sources) appear to be representative, those who agree to participate may not be; among respondents, young voters, for example, especially men, are likely to be underrepresented (Pennay et al. 2019: 52–53).

Since some groups will be underrepresented and others overrepresented, pollsters’ post-weight their samples. Which variables they used, neither Freshwater nor Resolve disclosed. Pollsters who did disclose, or did so subsequently, used at least two weights: gender (sex) and age (including age times sex). Most also weighted by education—ignored in 2019 (Pennay et al. 2019: 86)—region (including sex times region) and vote at the 2022 election (recalled). Half the polls were weighted by income (sometimes household, sometimes personal), one by housing tenure and one by religion (Table 20.2). How some of these were measured is another question.

These were not the only weights. Freshwater weighted its data by how respondents recalled voting in the 2023 Voice referendum, as did YouGov and Spectre (Table 20.2). According to Turner (2025), ‘“Labor-No” voters just didn’t switch over to the Coalition in the big numbers [we] estimated.’ But this may be only half the story. Polling conducted by The Australian National University between early 2025 and the election suggests that, while intending Labor voters who recalled voting ‘No’ were as likely to stay with Labor as those who recalled voting ‘Yes’, intending Coalition voters who said they had voted ‘Yes’ were 2.5 times more likely than ‘No’ voters to have moved to Labor (see Chapter 19, this volume). Were Freshwater’s figures different?

Table 20.2 Variables used to weight data, by poll, 2025

Poll

Sex

Age

Area

Education

Housing

Income

Religion

Vote 2022 election

Vote 2023 referendum

Roy Morgan

✓

✓

✓

–

–

–

–

✓

–

Freshwater*

–

–

–

–

–

–

–

–

–

Ipsos

✓

✓

✓

✓

–

household

–

✓

–

Newspoll

✓

✓

✓

✓

–

✓

–

✓

–

YouGov

✓

✓

✓

✓

–

household

–

✓

✓

DemosAU

✓

✓

✓

✓

✓

✓

–

✓

–

RedBridge/Accent

✓

✓

✓

✓

–

–

✓

–

–

Spectre

✓

✓

–

✓

–

✓

–

✓ State

✓

Resolve*

–

–

–

–

–

–

–

–

–

Essential

✓

✓

✓

✓

–

–

–

✓

–

Note: * Not disclosed.

Sources: Methodology disclosure statements from Australian Polling Council members (Ipsos, Pyxis, YouGov, Essential); personal communications (others).

The assumption behind weighting is that those who do not respond and those who do are the same in all relevant respects—‘ignorable’ nonresponse. But they may not be the same; hence, ‘“nonignorable” nonresponse’ (Bailey 2024: Ch. 5). Where respondents and non-respondents hold different views—when they are not the same in all relevant respects—weighting may compound the error. Those ‘who speak English poorly or not at all, those with low levels of literacy, and those with little digital affinity’ are likely to be underrepresented; so, too, are those with little interest in politics (Pennay et al. 2019: 53–57). Which of these is ‘ignorable’ and which ‘non-ignorable’ may be difficult to determine.

Post-survey errors may not be ignorable either. Those respondents classified as ‘undecided’ pose the most obvious problem. Respondents were classified as ‘undecided’ if they expressed no preference (Roy Morgan, RedBridge) or said ‘undecided’/‘don’t know’ even after being asked a ‘leaner’—a question to see whether there was a party (or candidate) towards which they were leaning even if it was not the party they were prepared to nominate when first asked how they would vote. Only Resolve and YouGov did not allow an ‘undecided’ response.

Table 20.3 Treatment of ‘undecided’ voters, by poll, 2025

Poll

Treatment

Roy Morgan

Like to see win

Freshwater

Not stated

Ipsos

Past vote

Newspoll

Excluded after leaner

YouGov

Not allowed

DemosAU

Excluded after leaner

RedBridge/Accent

Excluded after leaner

Spectre

Excluded after leaner

Resolve

Not allowed

Essential

Not allocated

Sources: AFR (2025); Beaumont (2025a); Butler (2025); Crowe (2025); DemosAU (2025); Ipsos Australia (2025); Pyxis Polling and Insights (2025); Roy Morgan (2025a); Spectre Strategy (2025); Wikimedia Foundation (2025); YouGov (2025b); Personal communications.

How did pollsters handle the ‘undecideds’? Where respondents did not respond positively to the ‘leaner’, DemosAU, Newspoll, RedBridge/Accent and Spectre dropped them—equivalent to assuming that they would divide in the same way as other respondents; that these three were among the polls that most accurately estimated the first preferences suggests that this was a reasonable move. Morgan, the other poll to score well on this measure, asked the ‘undecided’ which side they would like to see win—a question that ignored strategic voting and made sense only in relation to the 2PP vote. Ipsos, one of the least accurate, distributed the ‘undecideds’ according to how they said they had voted in 2022 (Table 20.3). Whether any of this mattered is unclear.

Single seats and clusters of seats

While most election-watchers may have been pondering the polls’ failure to predict the size of Labor’s victory, and searching for reasons, YouGov (2025c) declared that it ‘alone’ had ‘predict[ed] a historic win for Labor’ and was the only poll ‘to measure the scale of the Labor win’—a reference not to its final standard poll, conducted from 24 April to 1 May (Table 20.1), but to its very large MRP poll, completed two days earlier. Its previous (January–February) MRP, in sharp contrast, had suggested the Coalition might not only win but also do so without winning back any of the seats lost in 2022 to the Teals and the Greens (Goot 2025).

The enduring appeal of the MRP is twofold: its potential to overcome the limitations of national polling that does not use representative samples (Wang et al. 2015) and its promise to predict the outcome not only nationally but also seat by seat—Coalition seats, Labor seats and seats held by Independents and the Greens.

Based on its MRP, conducted 1–29 April, YouGov (2025a) estimated that both Labor and the Coalition would ‘get lower primary votes than at the last election’. The Coalition’s vote share, 31.8 per cent, came close to YouGov’s estimate of 31.1 per cent. Labor’s vote share, however, did not; it wasn’t 31.4 per cent, as YouGov estimated, but 34.6 per cent—two points higher than in 2022. From both its MRP and its standard polls, YouGov expected a 2PP swing to Labor (measured to the second decimal point) of 0.07 points. But it was based on its MRP, not its standard poll, that YouGov predicted a ‘Labor majority government’ (YouGov 2025a, 2025b).

The most likely result, the MRP reported, using a credibility interval of 90 per cent, was Labor winning 84 seats (an upper limit of 85, well short of the 94 seats that Labor actually won), the Coalition 47 (a lower limit of 45), Independents 14 (a lower limit of 13, not the 10 seats they did win) and the Greens three (a lower limit of two, not one). In short, the Labor majority YouGov envisaged was no higher than 20 seats; the majority Labor achieved, 38 seats, was nearly twice that.

How well did YouGov’s seat-by-seat projections fare? In the 124 seats it expected to be ‘classic contests’—in which the final two candidates in the count represented Labor and the Coalition (the Liberal Party, the Nationals or the LNP)—YouGov had Labor improving on its 2022 2PP vote in 69 but losing ground in 52 seats. In fact, Labor’s position improved in 102 of the 121 contests that were ‘classic’, losing ground in just 19 seats.

Among an anticipated 29 ‘non-classic’ contests—in which at least one of the final two candidates in the count did not represent a major party—there were 24 in which YouGov overestimated how well the Greens or Independent candidate would do, including 12 in which the overestimates exceeded 10 percentage points; in four, the contest turned out to be not ‘non-classic’ at all (Table 20.4).

Another way of judging an MRP is by calculating the absolute average error, seat by seat, in the 2PP/2CP (full tables available in Appendix 20.1). In the ‘marginal’ seats (seats requiring a swing of less than 6 percentage points), these errors were particularly high: a mean of 6.5 and a standard deviation of 7.54 in Coalition seats with a mean of 4.9 with a standard deviation of 3.77 in Labor seats. In ‘fairly safe’ seats for the Coalition, the figures are 3.7 and 5.37, respectively, and, for Labor, 5.1 and 4.94, respectively. In ‘safe’ Coalition seats, the figures are 3.0 and 3.16; in ‘safe’ Labor seats, 5.5 and 4.62, respectively.

One reason the MRP did not do well was because it was conducted over too long a period—out of the field close to election day having entered the field nearly five weeks earlier. As a comparison with the earlier MRP shows, and other polling confirms (see Chapter 19, this volume), Labor’s position improved substantially during the campaign. Being in the field for a month and not reporting on how opinion shifted over this time was a strategic error. Labor’s vote share was underestimated in 90 of the two-party contests (including 25 Coalition seats) and overestimated in just 34 (eight of them Labor seats). Since nominations were not declared until Friday, 11 April, when the MRP was launched, respondents could not be told candidates’ names or even presented with an official list of the parties contesting the seats and it appears that this did not change.

Table 20.4 Non-classic contests as expected by the AEC, YouGov or both: YouGov MRP, 1–29 April 2025 (per cent)

YouGov-predicted 2CP

Post-election 2CP

State/Territory

Electorate

Incumbent

AEC 2CP

ABC 2CP

Ind.

GRN

Coalition

ALP

Ind.

GRN

Coalition

ALP

Error

NSW

Bradfield

Lib.

53.4

52.5

64.1

–

35.9

–

50.0

–

50.0

–

14.1

Calare

Nats

59.7

59.7

53.1

–

46.9

–

56.8

–

43.2

–

–3.7

Cowper

Nats

52.4

52.4

63.2

–

36.8

–

47.5

–

52.5

–

15.7

Fowler

Ind.

51.8

51.1

63.0

–

37.0

–

52.7

–

–

47.3

10.3

Grayndler

ALP

64.0

67.3

–

34.7

–

65.3

–

33.1

–

66.9

1.6

Mackellar

Ind.

51.8

53.3

60.8

–

39.2

–

55.7

–

44.3

–

5.1

Richmond

ALP

58.2 (2PP)

58.2 (2PP)

–

47.4

–

52.6

–

–

40.0

60.0

–

Sydney

ALP

66.2

66.5

–

32.5

–

67.5

–

29.1

–

71.0

3.4

Warringah

Ind.

60.3

59.4

63.5

–

36.5

–

61.2

–

38.8

–

2.3

Wentworth

Ind.

50.6

56.8

58.4

–

41.6

–

58.3

–

41.7

–

0.1

Vic.

Cooper

ALP

58.9

57.8

–

39.8

–

60.2

–

40.3

–

59.7

–0.5

Goldstein

Ind.

51.8

53.3

63.5

–

36.5

–

49.9

–

50.1

–

13.6

Indi

Ind.

58.9

58.9

66.1

–

33.9

–

58.6

–

41.4

–

7.5

Kooyong

Ind.

52.5

52.2

60.1

–

39.9

–

50.7

–

49.3

–

9.4

Melbourne

GRN

55.6

56.5

–

56.1

–

43.9

–

47.0

–

53.0

9.1

Monash

Lib.

52.9 (2PP)

52.9 (2PP)

54.1

–

45.9

–

–

–

54.1

45.9

–

Nicholls

Nats

53.4

53.4

–

–

66.4

33.6

–

–

64.4

35.6

–

Wannon

Lib.

53.5

53.8

61.9

–

38.1

–

46.7

–

53.3

–

15.2

Wills

ALP

59.0

54.6

–

44.5

–

55.5

–

48.6

–

51.4

–4.1

Qld

Brisbane

GRN

53.7

53.7

–

–

41.4

58.6

–

–

41.0

59.0

–

Griffith

GRN

60.5

60.5

–

52.0

–

48.0

–

39.4

–

60.6

13.6

Groom

LNP

56.9

56.9

–

–

57.3

42.7

44.3

–

55.7

–

–

Kennedy

KAP

63.1

63.1

72.5

–

27.5

–

65.8

–

34.3

–

6.7

Maranoa

LNP

72.1 (2PP)

72.1 (2PP)

33.8

–

66.2

–

29.9

–

70.1

–

3.9

Ryan

GRN

52.7

52.6

–

52.0

48.0

–

–

53.3

46.7

–

–1.3

WA

Curtin

Ind.

51.3

51.3

55.7

–

44.3

–

53.3

–

46.7

–

2.4

SA

Mayo

CA

62.3

62.3

69.7

–

30.3

–

64.9

–

–

35.1

4.8

Tas.

Clark

Ind.

70.8

70.8

71.9

–

–

28.1

70.4

–

–

29.6

1.5

ACT

Bean

ALP

62.9 (2PP)

62.9 (2PP)

38.0

62.0

49.7

50.3

Canberra

ALP

62.2

62.2

–

39.9

–

60.1

–

30.5

–

69.5

9.4

Mean|median

6.6|4.6

Std dev.

5.97

Notes: CA = Centre Alliance; KAP = Katter’s Australian Party; Shaded result was predicted to be a non-classic contest but was a classic contest (AEC/ABC: Brisbane, Nicholls; YouGov: Richmond, Monash); or was predicted to be a classic contest but was a non-classic contest (Maranoa, Groom, Bean).

Source: ABC data from Green (2025).

Based on its MRP, YouGov expected 11 seats (10 Coalition, one Greens) to change hands. All but two did: Cowper and Wannon, predicted to go to Independents, stayed Liberal. Another eight also changed: seven from the Coalition to Labor (Hughes, NSW; Deakin, Victoria; Dickson, Forde and Petrie, Queensland) or to an Independent (Bradfield, NSW); one from the Greens to Labor (Melbourne, Victoria); and one from Independent to Liberal (Goldstein, Victoria). Of the 20 seats that changed, YouGov’s polling identified less than half.

In another poll—not an MRP—conducted 17–24 April, YouGov estimated vote shares in 10 classic regional contests and two non-classic contests. Again, results were calibrated to the second decimal point. In two of the classic contests (Dickson, where it was out by 11 points, and Bullwinkel, out by two), YouGov tipped the wrong winner. Across all 10 seats, the absolute error on 2PP averaged four points with a standard deviation of 4.53 (see Appendix 20.1 for seat-by-seat figures)—hardly an improvement on its MRP.

In other classic contests, polls also finished with large errors (10 points or more): KJC in Hunter, Richmond, Blair and Tangney, though the real size of its errors are difficult to judge; UComms and Freshwater in Dickson, with the Freshwater poll said to have misled the Liberal Party into thinking that its leader, Peter Dutton, was not in danger; DemosAU in Adelaide (a safe seat misdescribed as marginal); UComms, in Lyons, and in Forrest, where it failed to anticipate who the last two candidates would be; the New England Times, in New England; and a Compass poll in McMahon, which had Labor’s Chris Bowen (who went on to win 45.5 per cent of the first-preference vote) losing to an Independent, Matthew Camenzuli (who secured just 9.8 per cent of the vote)—with Camenzuli listed first on the ballot (Bowen’s spot) with a tick next to his name, but the APC reluctant to condemn it (ABC 2025). The only poll able to withstand similar scrutiny was the EMRS in Franklin.

Non-classic contests also threw up some ‘stinkers’ (Beaumont 2025b). In Bradfield, YouGov underestimated the Liberals’ two-candidate vote by 14.1 points, in Wannon by 15.2 and in Cowper by 15.7 points. In Ryan, retained by the Greens, JWS forecast a classic contest; in Griffith, where Labor beat the Greens 60.6–39.4, JWS had Labor leading the LNP 51–49, as though it were also a classic contest; and in Brisbane, where JWS had Labor ahead 51–49 against the LNP, Labor prevailed 59–41 per cent.

Table 20.5 Contests in clusters of seats, 31 March – 3 May 2025 (per cent)

First preferences

2PP (ALP)

Pollster

Dates

Electorates

n

ALP

Lib./Nats

GRN

PHON

Other

DemosAU

7–15 April

Brisbane, Griffith, Ryan (Qld) [held by Greens]

1,087

Est.

29.0

36.0

29.0

3.0

3.0

56.0

Vote

31.7

31.8

28.8

2.5

5.2

n.a.

Error

–2.7

+4.2

+0.2

+0.5

–2.2

DemosAU

13–22 April

Bruce, Dunkley, Hawke (Vic.) [Melbourne Labor marginals]

924

Est.

32.0

31.0

13.0

10.0

14.0

53.0

Vote

40.9

28.5

11.2

8.2

11.2

59.8

Error

–8.9

+2.5

+1.8

+1.8

+2.8

–6.8

DemosAU

18–23 April

Bonner, Dickson, Forde, Longman, Petrie (Qld) [Brisbane LNP outer metropolitan]

1,053

Est.

27.0

40.0

13.0

7.0

13.0

47.0

Vote

35.8

35.0

10.6

7.1

13.3

52.6

Error

–8.8

+5.0

+2.4

–0.1

–0.3

–5.6

DemosAU

13–27 April

Parramatta, Reid, Werriwa (NSW) [Sydney Labor marginals]

905

Est.

36.0

28.0

10.0

4.0

22.0

56.0

Vote

45.9

31.1

11.6

3.0

8.4

60.6

Error

–9.9

+3.1

–1.6

+1.0

+13.6

–4.6

RedBridge/Accent

24–30 April

Bennelong, Dobell, Gilmore, Lindsay, Paterson, Robertson, Werriwa (NSW); Aston, Casey, Chisholm, Corangamite, Menzies (Vic.); Blair, Forde, Leichhardt, Longman (Qld); Bullwinkel, Tangney (WA); Sturt (SA); Lyons (Tas.) [20 ‘key seats’: 13 Labor, 7 Coalition]

1,004

Est.

33.0

34.0

12.0

6.0

16.0

53.0

Vote

37.3

33.8

10.9

5.9

12.1

55.7

Error

–4.3

+0.2

+1.1

+0.1

+3.9

–2.7

Sources: Publicly reported estimates; AEC (2025a); author’s calculations.

One of the things that JWS got wrong in these seats—and in others, such as Whitlam and Werriwa, which were not fully reported—were preference flows. Some 80–90 per cent of the preferences from the small right-wing parties (One Nation, Family First, Trumpet of Patriots) were likely to flow to the Coalition, it said, not the two-thirds or so that flowed in 2022 (Coorey 2025a).

In other polls, seats were clustered. DemosAU polled the three Brisbane seats held by the Greens (7–15 April), where its estimate of the 2CP vote assumed either an overall Labor versus LNP or Greens versus LNP contest (in Griffith, it turned out to be Greens versus Labor); three Labor marginals in Melbourne (13–22 April); five outer metropolitan seats in Queensland held by the LNP (18–23 April); and three Labor marginals in Sydney (13–27 April). In every case, the final Labor vote was underestimated—its first-preference vote more markedly than its 2PP share, with the 2PP share more accurate the closer the poll was to election day (Table 20.5). There was also a RedBridge/Accent poll of 20 ‘key seats’—13 Labor, seven Coalition—conducted six times, most recently on 24–30 April; all but three were marginals. It, too, underestimated Labor’s first preferences and (more narrowly) Labor’s share of the 2PP vote.

National polls versus the seat-by-seat count in the MRP

To focus on national polls rather than polls in particular seats is to assume a close relationship between changes in total vote share and in seat shares: seats won by the parties of government—hence, the electoral pendulum (Green 2024). As Phil Coorey (2025b) observed, if the 2PP ‘swing against Labor of 0.6 percentage points’—the swing predicted by the Freshwater poll—‘was applied across all electorates’, Labor would be reduced from 78 seats to ‘a bare majority’ of 76, with the Coalition increasing its share from 57 to 59 seats. For this to be valid, swings did not need to be uniform; all that was required was that seats lost despite requiring a greater than average swing were matched by seats gained despite requiring a smaller than average swing.

Given the rise of the Independents, and the growing view that the pendulum had passed its use-by date, the performance of the pendulum was remarkable. A swing of 3.2 points might have been expected to deliver 13 seats from the Coalition to Labor (derived from Green 2024). In fact, the swing delivered 11 seats to Labor from the Coalition, just two seats short. (Aston, counted by the AEC [2025b] as a Labor gain, was Labor already, following the 2023 by-election; Bennelong, notionally a Labor gain, was already regarded as Labor after the 2024 redistribution.)

By contrast, the MRP fared poorly. It predicted a 2PP vote of 52.9 per cent—a swing to Labor of just 0.8 of a point. On the pendulum, this might have been expected to have shifted three of the Coalition’s seats to Labor. The MRP’s seat-by-seat predictions—the rationale for conducting an MRP—had Labor gaining seven seats from the Coalition, six of them marginals, with the other (Braddon) ‘fairly safe’. All seven did fall to Labor, but another six fell as well.

Conclusion

The national polls exaggerated the Coalition’s position (and One Nation’s) and underestimated Labor’s. This was true for most of the other polls as well. As in 2019, when the average error in the national polls was the same, 2025 was not a ‘miss’, explained by a late swing, but a ‘failure’, due to pollsters’ errors (Durand and Blais 2020). Misses are uncommon, failures are not (Goot 2021). Had the polls predicted not a Labor majority but a Coalition victory or even a hung parliament, their performance would have generated more concern. In 2019, when the polls picked the wrong winner, the Market Research Society commissioned an inquiry (Pennay et al. 2019); after the 2025 election, because no poll suggested that the Coalition would win, no-one called for an inquiry.

Possible explanations for the polls’ failure include survey errors (notably, non-ignorable nonresponses) and post-survey errors, including the treatment of the ‘undecideds’. Might herding have played a part—even the dominant part—with pollsters making sure their poll was not too different from that of their rivals, being in wrong company preferable to being out on one’s own? After the 2019 debacle, one pollster said he had not published his figures because they were at odds with those published by others (Mansillo and Jackman 2020: 137); what his figures were, he never disclosed.

In 2025, it is more likely that pollsters simply subscribed to a common assumption—widely shared by politicians and political journalists—that, in a cost-of-living ‘crisis’, if Labor were to win, it would not be by much. This proved mistaken: Labor won in a landslide.

References

Australian Broadcasting Corporation (ABC). 2025. ‘Poll position.’ Media Watch, [ABC TV], 28 April. www.abc.net.au/mediawatch/episodes/poll/105225644.

Australian Electoral Commission (AEC). 2025a. ‘House of Representatives—Final results.’ [Updated 10 June 2025]. Tally Room: 2025 Federal Election. Canberra: Australian Electoral Commission. results.aec.gov.au/31496/Website/HouseDefault-31496.htm.

Australian Electoral Commission (AEC). 2025b. ‘Seats that changed hands.’ [Updated 10 June 2025]. Tally Room: 2025 Federal Election. Canberra: Australian Electoral Commission. results.aec.gov.au/31496/Website/HouseSeatsWhichChangedHands-31496.htm.

Australian Financial Review (AFR). 2025. ‘Explore the latest Freshwater strategy poll.’ Australian Financial Review, 23 February. [Updated 1 May 2025]. www.afr.com/politics/federal/freshwater-strategy-poll-20250128-p5l7p3.

Australian Polling Council. 2025. Code of Conduct. [Online]. www.australianpollingcouncil.com/code-of-conduct.

Bailey, Michael A. 2024. Polling at a Crossroads: Rethinking Modern Survey Research. Cambridge: Cambridge University Press. doi.org/10.1017/9781108697798.

Beaumont, Adrian. 2025a. ‘Major YouGov poll has Labor easily winning a majority of seats in election.’ The Conversation, 1 May. doi.org/10.64628/AA.puwm55w3g.

Beaumont, Adrian. 2025b. ‘Final counting shows polls underrated Labor in 2025 election almost as much as they overstated it in 2019.’ The Conversation, 5 June. doi.org/10.64628/AA.cpjyxwpwu.

Beed, Terence W. 1977. ‘Opinion polling and the elections.’ In Australia at the Polls: The National Elections of 1975, edited by Howard R. Penniman, 211–55. Canberra: Australian National University Press.

Bowe, William. 2025a. ‘Federal election plus one week (open thread).’ [Online]. The Poll Bludger, 11 May. www.pollbludger.net/2025/05/11/federal-election-plus-one-week-open-thread/.

Bowe, William. 2025b. ‘Federal politics: Roy Morgan poll and preference flow data (open thread).’ [Online]. The Poll Bludger, 25 June. www.pollbludger.net/2025/06/25/federal-politics-roy-morgan-poll-and-preference-flow-data-open-thread/.

Butler, Josh. 2025. ‘Guardian Essential Poll: Labor leads Coalition in final pre-election poll as Dutton’s approval rating slips further.’ The Guardian, 29 April. www.theguardian.com/australia-news/2025/apr/29/guardian-essential-poll-labor-leads-coalition-in-final-pre-election-poll-as-duttons-approval-rating-slips-further.

Coorey, Phillip. 2025a. ‘One Nation gives Dutton a poll boost.’ Australian Financial Review, 30 April.

Coorey, Phillip. 2025b. ‘Final poll: Labor fighting to retain majority government.’ Australian Financial Review, 1 May. www.afr.com/politics/federal/labor-fighting-to-retain-majority-government-20250501-p5lvkp.

Crowe, David. 2025. ‘Labor holds clear lead as time runs short for a Coalition rebound, poll reveals.’ The Sydney Morning Herald, 29 April. www.smh.com.au/politics/federal/labor-holds-its-lead-as-time-runs-short-for-a-liberal-rebound-poll-reveals-20250429-p5luz8.html.

DemosAU. 2025. Final Poll: 2025 Australian Election. [Report]. 2 May. demosau.com/wp-content/uploads/2025/05/DemosAU-Report-Final-Poll-May-2025-Australian-Election-1.pdf.

Durand, Claire, and André Blais. 2020. ‘Quebec 2018: A failure of the polls?’ Canadian Journal of Political Science 53, no. 1: 133–50. doi.org/10.1017/S0008423919000787.

Goot, Murray. 2021. ‘How good are the polls? Australian election predictions, 1993–2019.’ Australian Journal of Political Science 56, no. 1: 35–55. doi.org/10.1080/10361146.2020.1825616.

Goot, Murray. 2025. ‘A poll that answers Dutton’s dreams …’. Inside Story, 10 March. insidestory.org.au/a-poll-that-answers-duttons-dreams/.

Green, Antony. 2024. 2025 Federal Election—New Electoral Pendulum. Version 1a, December 2024. Sydney: ABC Election Unit. www.abc.net.au/dat/news/elections/federal/2025/guide/FED2025_PostRedistPendulum.pdf.

Green, Antony. 2025. ‘FED2025—Tracking the early vote.’ Antony Green’s Election Blog, 27 April. antonygreen.com.au/fed2025-tracking-the-early-vote/.

Ipsos Australia. 2025. ‘Long methodology disclosure statement: Ipsos Australian national poll.’ Media release, 2 May. Sydney: Ipsos Australia. www.ipsos.com/sites/default/files/2025-05/Ipsos%20-%20APC%20Long%20Methodology%20Disclosure%20Statement_Poll%20%233.pdf.

Jennings, Will, and Christopher Wlezien. 2018. ‘Election polling errors across time and space.’ Nature Human Behaviour 2: 276–83. doi.org/10.1038/s41562-018-0315-6.

Knott, Matthew. 2025. ‘Pollsters left red-faced by Labor’s powerful late surge.’ The Sydney Morning Herald, 6 May.

Mansillo, Luke, and Simon Jackman. 2020. ‘National polling and other disasters.’ In Morrison’s Miracle: The 2019 Australian Federal Election, edited by Anika Gauja, Marian Sawer, and Marian Simms, 125–48. Canberra: ANU Press. doi.org/10.22459/MM.2020.07.

Pennay, Darren, Murray Goot, Dina Neiger, Dennis Trewin, Paul J. Lavrakas, John Stirton, Phil Hughes, Jill Sheppard, and Ian McAllister. 2019. Report of the Inquiry into the Performance of the Opinion Polls at the 2019 Australian Federal Election. Association of Market and Social Research Organisations and the Statistical Society of Australia.

Pyxis Polling and Insights. 2025. ‘Methodology statement: Newspoll.’ Media release, 2 May. Sydney: Pyxis Polling and Insights. www.pyxispolling.com/ws/media-library/906f8ff5e6c511d2ba153b238bc19646/newspoll-methodology-statement-2-may-2025-by-pyxis.pdf.

Read, Michael. 2025. ‘This is why some pollsters missed the Labor landslide.’ Australian Financial Review, 4 May. www.afr.com/politics/this-is-why-some-pollsters-missed-the-labor-landslide-20250504-p5lwet.

Reed, Jim. 2025. ‘With 77% of the vote now counted, we thought it was time to run an initial analysis of the polling data.’ LinkedIn post. www.linkedin.com/feed/update/urn:li:activity:7324957787187269633/.

Roy Morgan. 2025a. ‘ALP set to win federal election with an expanded majority—ALP 53% cf. L-NP 47%.’ Article no. 9886, 2 May. Melbourne: Roy Morgan. roymorgan-cms-prod.s3.ap-southeast-2.amazonaws.com/wp-content/uploads/2025/05/02093140/9886-Federal-Voting-Intention-Final-Pre-Election-May-2-2025.pdf.

Roy Morgan. 2025b. ‘Roy Morgan update May 6, 2025: Roy Morgan aces election, consumer confidence and business confidence.’ Finding no. 9869, Press release, 6 May. Melbourne: Roy Morgan. www.roymorgan.com/findings/roy-morgan-update-may-6-2025.

Spectre Strategy. 2025. Australian Election Vote Forecast 2025. [Online]. 1 May. www.spectrestrategy.com/blog-3-1/auselectionvoteforecast.

Turner, Michael. 2025. ‘Three reasons we missed Labor’s landslide.’ Australian Financial Review, 4 May. www.afr.com/politics/federal/three-reasons-we-missed-labor-s-landslide-20250504-p5lwf8.

Wang, Wei, David Rothschild, Sharad Goel, and Andrew Gelman. 2015. ‘Forecasting elections with non-representative polls.’ International Journal of Forecasting 31, no. 3: 980–91. doi.org/10.1016/j.ijforecast.2014.06.001.

Wikimedia Foundation. 2025. ‘Opinion polling for the 2025 Australian federal election.’ Wikipedia, [Updated 19 August 2025]. en.wikipedia.org/w/index.php?title=Opinion_polling_for_the_2025_Australian_federal_election&oldid=1306694375.

YouGov. 2025a. ‘Labor to win with an increased majority in YouGov’s final MRP of the election.’ YouGov Australia, 1 May. au.yougov.com/politics/articles/52098-labour-to-win-with-an-increased-majority-in-yougovs-final-mrp-of-the-election.

YouGov. 2025b. ‘Labor on track for victory as Dutton records lowest satisfaction rating on record.’ YouGov Australia, 2 May. au.yougov.com/politics/articles/52103-labor-on-track-for-victory-as-dutton-records-lowest-satisfaction-rating-on-record.

YouGov. 2025c. ‘YouGov accurately calls historic Labor landslide in Australian federal election.’ YouGov Australia, 7 May. au.yougov.com/politics/articles/52115-yougov-accurately-calls-historic-labor-landslide-in-australian-federal-election.

Appendix 20.1

Table A20.1 The MRP poll in ‘marginal’ seats, by order of marginality: YouGov, 1–29 April 2025 (per cent)

Coalition 2PP/2CP

Labor 2PP/2CP

Electorate

State

Margin

Predicted

Actual

Error

Electorate

State/Territory

Margin

Predicted

Actual

Error

Deakina

Vic.

0.0

49.3

47.2

+2.1

Gilmore

NSW

0.2

51.2

55.1

–3.9

Bennelong

NSW

0.0

49.0

40.7

+8.3

Menzies

Vic.

0.4

50.3

51.1

–0.8

Sturt

SA

0.5

49.3

43.4

+5.9

Lyons

Tas.

0.9

52.3

61.6

–9.3

Moore

WA

0.9

49.7

47.1

+2.6

Lingiari

NT

1.7

58.0

58.1

–0.1

Canning

WA

1.2

50.2

56.6

–6.4

Robertson

NSW

2.2

50.9

59.4

–8.5

Bassa

Tas.

1.4

52.2

42.0

+9.8

Paterson

NSW

2.6

51.8

56.9

–4.9

Casey

Vic.

1.4

52.7

52.9

–0.2

Tangney

WA

2.8

53.5

57.0

–3.5

Dicksona

Qld

1.7

50.2

44.0

+8.2

Boothby

SA

3.3

51.9

61.1

–9.2

Cowperb vs Ind.

NSW

2.4

36.8

52.5

–15.7

Chisholm

Vic.

3.3

53.1

55.7

–4.6

Bradfield vs Ind.

NSW

2.5

35.9

50.0

–14.1

Bullwinkel

WA

3.3

54.1

50.5

+3.6

Banks

NSW

2.6

49.4

47.6

+1.8

Astonb

Vic.

3.6

48.0

53.4

–5.4

Monash

Vic.

2.9

45.9

54.1

–8.2

Parramatta

NSW

3.7

56.3

62.6

–6.3

Longman

Qld

3.1

54.2

50.1

+4.1

McEwen

Vic.

3.8

53.2

55.8

–2.6

Bonner

Qld

3.4

48.8

45.0

+3.8

Wills vs GRN

Vic.

4.6

55.5

51.4

+4.1

Leichhardt

Qld

3.4

52.8

43.9

+8.9

Hunter

NSW

4.8

55.0

59.5

–4.5

Nicholls

Vic.

3.4

66.4

64.4

+2.0

Reid

NSW

5.2

57.4

62.0

–4.6

Hughesa

NSW

3.5

54.6

46.9

+7.7

Blair

Qld

5.2

54.5

55.7

–3.2

Flynn

Qld

3.8

52.8

60.2

–7.4

Bruce

Vic.

5.3

55.2

64.6

–9.4

Wannon vs Ind.

Vic.

3.8

38.1

53.3

–15.2

Werriwa

NSW

5.3

53.3

56.8

–3.5

Forrest

WA

4.2

53.9

52.2

+1.7

Mean|median

4.9|–4.6

Fordea

Qld

4.2

52.5

48.2

+4.3

Std dev.

3.8

Petriea

Qld

4.4

53.4

48.8

+4.6

Durack

WA

4.7

54.9

60.2

–5.3

Bowman

Qld

5.5

51.4

52.4

–1.0

Mean|median

5.5|+2.1

Std dev.

7.5

Notes: Marginal = seats requiring a 2PP swing of less than 6 percentage points to change hands. Error = predicted minus actual. Seats in bold text changed hands; a Seat changed hands but not predicted to change hands; b Seat predicted to change hands but did not change hands.

Sources: Green (2024); YouGov (2025a); AEC (2025a); author’s calculations.

Table A20.2 The MRP poll in ‘fairly safe’ seats, by order of marginality: YouGov, 1–29 April 2025 (per cent)

Coalition 2PP

Labor 2PP

Electorate

State

Margin

Predicted

Actual

Error

Electorate

State/Territory

Margin

Predicted

Actual

Error

Lindsay

NSW

6.1

55.1

52.8

+2.3

Shortland

NSW

6.0

55.6

61.5

–3.9

Flinders

Vic.

6.2

54.9

54.8

+0.1

Eden-Monaro

NSW

6.1

54.9

57.2

–2.3

Capricornia

Qld

6.6

55.8

55.8

0.0

Macquarie

NSW

6.3

54.9

57.7

–2.8

O’Connor

WA

6.7

55.1

63.3

–8.2

Dobell

NSW

6.6

55.5

59.4

–3.9

Hume

NSW

6.9

59.2

58.1

+1.1

Dunkley

Vic.

6.8

57.5

57.1

+0.4

Groom vs Ind.

Qld

6.9

57.3

63.4

–6.1

Holt

Vic.

7.1

55.1

64.0

–8.9

Berowra

NSW

7.6

53.8

51.6

+2.2

Hawke

Vic.

7.6

54.1

57.6

–3.5

Braddona

Tas.

8.0

57.0

42.8

+14.2

Cooper

Vic.

7.8

60.2

78.5

–18.3

La Trobe

Vic.

8.4

56.3

52.1

+4.2

Corangamite

Vic.

7.8

57.5

58.1

–0.6

Fisher

Qld

8.7

55.3

56.0

–0.7

Greenway

NSW

7.9

59.4

63.8

–4.4

Fairfax

Qld

9.0

55.8

53.2

+2.6

Richmond

NSW

8.2

52.6

60.0

–7.4

McPherson

Qld

9.3

55.1

54.4

+0.7

Whitlam

NSW

8.3

59.2

56.3

+2.9

Riverina

NSW

9.7

57.4

62.6

–5.2

Solomon

NT

8.4

59.7

51.3

+8.4

Calare vs Ind. MP

NSW

9.7

46.9

43.2

+3.7

Pearce

WA

8.8

56.0

56.4

–0.4

Mean|median

3.7|0.9

Hindmarsh

SA

8.9

60.4

66.4

–6.0

Std dev.

5.37

Rankin

Qld

9.1

59.4

65.6

–6.2

Moreton

Qld

9.1

59.9

66.1

–6.2

Swan

WA

9.4

58.5

64.0

–5.5

Isaacs

Vic.

9.5

60.0

64.3

–4.3

Macarthur

NSW

9.8

59.1

65.6

–6.5

Cowan

WA

9.9

59.4

63.6

–4.2

Mean|median

5.1|4.2

Std dev.

4.94

Notes: Fairly safe = seats requiring a 2PP swing of between 6.0 and 9.9 percentage points to change hands; Error = predicted less actual; Seats in bold text changed hands; a Seat changed hands but not predicted to change hands.

Sources: Green (2024); YouGov (2025a); AEC (2025a); author’s calculations.

Table A20.3 The MRP poll in ‘safe’ seats, by order of marginality: YouGov, 1–29 April 2025 (per cent)

Coalition 2PP

Labor 2PP

Electorate

State

Margin

Predicted

Actual

Error

Electorate

State/Territory

Margin

Predicted

Actual

Error

Grey

SA

10.1

57.2

54.6

+2.6

Gorton

Vic.

10.0

56.2

60.3

–4.1

Hinkler

Qld

10.1

56.2

56.3

–0.1

Hasluck

WA

10.0

66.0

66.0

0.0

Dawson

Qld

10.4

59.0

61.8

–2.8

McMahon

NSW

10.5

60.8

59.0

+1.8

Mitchell

NSW

10.5

56.5

53.8

+2.7

Lilley

Qld

10.5

62.5

64.5

–2.0

Fadden

Qld

10.6

55.9

56.9

–1.0

Makin

SA

10.8

61.5

64.7

–3.2

Page

NSW

10.7

58.0

59.3

–1.3

Bendigo

Vic.

11.2

60.9

51.4

+9.5

Wright

Qld

10.9

58.6

58.0

+0.6

Gellibrand

Vic.

11.2

62.3

65.1

–2.8

Moncrieff

Qld

11.2

58.1

58.8

–0.7

Hotham

Vic.

11.6

63.6

66.9

–3.3

Wide Bay

Qld

11.3

54.5

57.6

–3.1

Oxley

Qld

11.6

62.7

69.2

–6.5

Cook

NSW

11.6

59.0

57.2

+1.8

Adelaide

SA

11.9

61.9

69.1

–7.2

Herbert

Qld

11.8

59.4

63.4

–4.0

Barton

NSW

12.0

64.1

66.0

–1.9

Lyne

NSW

13.8

54.9

59.8

–4.9

Macnamara

Vic.

12.2

58.0

61.8

–3.8

New England

NSW

15.2

61.2

67.1

–5.9

Jagajaga

Vic.

12.2

60.8

62.9

–2.1

Farrer

NSW

16.3

59.3

62.9

–3.6

Canberra vs GRN

ACT

12.2

60.1

76.4

–15.3

Barker

SA

16.6

60.7

63.0

–2.3

Calwell

Vic.

12.4

58.5

64.7

–6.2

Parkes

NSW

18.2

62.7

63.0

–0.3

Corio

Vic.

12.5

60.2

63.2

–3.0

Mallee

Vic

19.0

64.6

69.0

–4.4

Lalor

Vic.

12.8

59.5

63.2

–3.7

Gippsland

Vic.

20.6

61.5

69.4

–7.9

Spence

SA

12.9

59.9

65.3

–5.4

Maranoa vs PHON

Qld

22.1

66.2

74.0

–7.8

Bean

ACT

12.9

62.0

69.3

–7.3

Mean|median

3.0|–2.3

Ballarat

Vic.

13.0

61.6

60.7

+0.9

Std dev.

3.16

Maribyrnong

Vic.

13.0

63.5

62.7

+0.8

Blaxland

NSW

13.0

63.4

71.9

–8.5

Burt

WA

13.3

63.3

65.7

–2.4

Kingsford Smith

NSW

13.3

65.2

67.2

–2.0

Chifley

NSW

13.6

62.3

69.8

–7.5

Franklin

Tas.

13.7

58.6

69.4

–10.8

Perth

WA

14.4

56.6

66.5

–9.9

Cunningham

NSW

15.1

62.2

67.5

–5.3

Watson

NSW

15.2

64.7

72.9

–8.2

Scullin

Vic.

15.4

62.1

64.3

–2.2

Fenner

ACT

15.7

66.6

72.1

–5.5

Kingston

SA

16.4

64.2

70.7

–6.5

Sydney vs GRN

NSW

16.5

67.5

78.1

–10.6

Fraser

Vic.

16.6

63.3

72.0

–8.7

Fremantle

WA

16.9

62.2

68.7

–6.5

Brand

WA

17.1

65.9

66.9

–1.0

Grayndler vs GRN

NSW

17.3

65.3

80.2

–14.9

Newcastle

NSW

17.9

64.7

70.8

–6.1

Mean|median

5.5|–4.4

Std dev.

4.62

Note: Safe = seats requiring a 2PP swing of more than 10 percentage points to change hands.

Sources: Green (2024); YouGov (2025a); AEC (2025a); author’s calculations.

Table A20.4 Classic contests in single electorates, 31 March – 3 May 2025 (per cent)

First preferences

Electorate

Incumbent; margin

Pollster

Date(s)

n

ALP

Coalition

GRN

PHON

Ind.

Other

2PP (ALP)

NSW

Eden-Monaro

ALP (McBain); 6.1

YouGov

17–24 April

268

Est.

42.2

26.0

10.9

7.0

10.6

3.3

61.0

–

Vote

43.0

31.9

10.0

7.0

3.9

4.2

57.2

Gilmore

ALP (Phillips); 0.2

YouGov

17–24 April

255

Est.

36.2

33.5

11.1

6.0

8.5

4.7

54.0

–

Vote

38.1

34.5

7.1

5.0

7.5

7.8

55.1

Hunter

ALP (Repacholi); 4.8

KJC Research

20 April

600

Est.

–

–

–

–

–

–

45.0*

YouGov

17–24 April

224

Est.

35.8

14.6

8.8

25.3

0.5

15.0

59.0

–

Vote

43.5

18.2

7.4

16.1

–

14.8

59.5

McMahon

ALP (Bowen); 10.5

Compass

8 April

1,003

Est.

19.0

20.0

–

–

41.0

–

–

–

Vote

45.5

26.8

9.1

8.7

9.8

–

59.0

New England

Nats (Joyce); 15.2

New England Times

9–23 April

426

Est.

24.1

43.6

12.2

6.8

2.4

3.8

42.7

–

Vote

20.3

52.2

7.9

10.0

3.7

5.9

32.9

Paterson

ALP (Swanson); 2.6

YouGov

17–24 April

302

Est.

30.9

25.9

11.5

14.2

15.9

1.6

55.0

–

Vote

37.0

27.2

7.6

7.6

12.1

8.4

56.9

Richmond

ALP (Elliot); 8.2

KJC Research

24 April

600

Est.

–

–

–

–

–

–

34.0*

–

Vote

30.4

24.6

26.5

5.4

2.6

10.5

60.0

Vic.

Ballarat

ALP (King); 13.0

YouGov

17–24 April

217

Est.

38.5

23.8

17.7

9.8

6.1

4.1

62.0

–

Vote

42.3

28.6

14.3

7.7

2.8

4.3

60.7

Qld

Blair

ALP (Neumann); 5.2

KJC Research

24 April

600

Est.

–

–

–

–

–

–

41.0*

–

Vote

–

–

–

–

–

–

55.7

Dickson

LNP (Dutton); 1.7

Freshwater

10 April (pub.)

–

Est.

–

–

–

–

–

–

43.0

UComms

9–10 April

854

Est.

24.2

37.6

10.9

–

12.0

4.6

52.0

YouGov

17–24 April

253

Est.

24.2

40.3

7.6

5.4

16.5

6.0

45.0

–

Vote

33.6

34.8

7.6

4.2

12.2

7.6

56.0

WA

Bullwinkel

ALP (notional); 3.3

YouGov

17–24 April

238

Est.

31.6

40.6

11.1

6.1

0.6

6.6

48.5

–

Vote

32.0

40.1

11.2

8.6

–

8.2

50.5

Forrest

Liberal; 4.2

UComms

9–10 April

c.1,000

Est.

–

34

–

–

20

–

49.0 (Lib.)

–

Vote

22.6

36.0

7.9

8.7

18.3

6.6

52.2 (Lib.)

Tangney

ALP (Lim); 2.8

KJC Research

24 April

600

Est.

–

–

–

–

–

–

45.0*

–

Vote

42.5

34.3

13.0

4.1

–

6.2

57.0

SA

Adelaide

ALP (Georganas); 11.9

DemosAU

27–29 April

1,974

Est.

29

32

12

9

–

18

51

–

Vote

46.5

24.2

19.0

4.0

–

6.4

69.1

Boothby

ALP (Miller-Frost); 3.3

YouGov

17–24 April

250

Est.

32.1

30.7

14.9

9.0

0.4

12.9

52

–

Vote

42.6

32.5

17.1

3.0

–

4.8

61.1

Tas.

Braddon

Liberal; 8.0

YouGov

17–24 April

419

Est.

33.2

30.6

9.7

4.6

15.7

6.2

54

–

Vote

39.5

31.7

8.4

7.6

8.3

4.5

57.2

Franklin

ALP (Collins); 13.7

EMRS

9–10 April

430

Est.

38

19

13

–

20

–

–

–

Vote

39.0

18.8

10.5

5.0

26.7

–

57.8

Lyons

ALP; 0.9

UComms

10 April (pub.)

712

Est.

27.2

29.5

14.6

4.1

–

5.8

50.9

YouGov

17–24 April

446

Est.

33.4

27.8

13.0

4.4

9.3

12.1

56

–

Vote

43.3

26.4

10.6

6.8

3.2

9.7

61.6

Note: * Two-party vote adds to less than 100.

Sources: Publicly reported estimates; YouGov, Personal communication; AEC (2025a).

Table A20.5 Non-classic contests in single seats, 31 March – 3 May 2025 (per cent)

First preferences

Electorate

Incumbent; margin

Pollster

Date(s)

n

ALP

Coalition

GRN

PHON

Ind.

Other

2CP (Ind.)

NSW

Calare

Gee (Ind.); 9.7 vs Nats

YouGov

17–24 April

184

Est.

8.2

25.1

5.1

13.9

38.9

8.8

57.0

–

Vote

10.4

29.7

3.5

7.7

39.4

9.3

56.8

Wentworth

Spender (Ind.); 6.8 vs Lib.

UComms

9 April (pub.)

1,015

Est.

12.2

32.9

–

–

32.5

–

58.0

Compass

18 April (pub.)

627

Est.

15.0

47.0

10.0

–

28.0

–

–

–

Vote

13.4

36.4

10.2

2.4

37.7

–

58.3

Vic

Wannon

Liberal (Tehan); 3.8 vs Ind.

YouGov

17–24 April

246

Est.

12.2

33.2

5.2

6.9

35.7

6.8

57.0

–

Vote

10.6

43.7

3.1

4.2

32.1

6.3

46.3

Qld

Brisbane

Greens (Bates); 3.7 vs LNP

JWS Research

8–9 April

834

Est.

29.0

43.0

29.0

–

–

–

51 (ALP)

–

Vote

32.2

34.3

25.9

2.5

–

5.2

59.0 (ALP)

Griffith

Greens (Chandler-Mather); 0.9 vs ALP (primary)

JWS Research

8–9 April

888

Est.

28.0

38.0

23.0

–

–

–

51(ALP 2PP)

–

Vote

34.5

26.6

31.7

2.5

–

4.8

60.6 (ALP)

Ryan

Greens (Watson-Brown); 2.6 vs LNP

JWS Research

8–9 April

825

Est.

30.0

46.0

13.0

–

–

11.0

57 (LNP 2PP)

–

Vote

28.2

34.6

29.0

2.3

–

5.9

53.3 (GRN)

Sources: Publicly reported estimates; JWS, Personal communication; AEC (2025a); Green (2024).


  1. 1 For additional information, I thank Jessica Elgood from Ipsos, George Hasanakos from DemosAU, Morgan James from Spectre, Julian McCrann from Roy Morgan, Shaun Ratcliff from Accent, Kos Samaras from RedBridge, John Scales from JWS, Paul Smith from YouGov, John Utting from Utting Research, Dorian von Freyhold from Compass and Campbell White from Pyxis. Freshwater did not respond to my requests and Resolve asked me not to make any. For their comments on an earlier draft, I thank Anika Gauja, Luke Mansillo and Jill Sheppard.


Previous

Updated:  11 August 2026/ Responsible Officer:  / Page Contact: