How wrong have Australian polls been? Primaries, the 2PP and the λ test
Abstract. At the seven House elections from 2007 to 2025, the average of the final polls missed each main party's primary vote by 1.4 points. Labor was overstated at five, but in 2025 every final poll had Labor too low. The pollsters' own 2PP missed by 1.6 points on average with no steady direction. We define the error, describe the 46 final polls and their sources, report the gap between the two largest parties, each pollster's record and state elections as a check, and show why the leave-one-out test rejects a correction.
Summary
This paper sets out how we measure polling error in Australia: which polls count as final polls, where they come from, how the error is defined on the primary vote and on the two-party-preferred (2PP), and what the record shows. At the seven House elections from 2007 to 2025 for which final primaries survive, the average of the final polls missed each main party's primary vote by 1.4 points. Labor was overstated at five of the seven, but the exception, 2025, was the largest Labor miss in the record: every one of the ten final polls had Labor too low, by between 0.6 and 5.1 points. On the 2PP, the polls' famous miss of 2019 (Labor ahead, Labor lost) was matched in 2025 by a miss of the same size the other way.
The record has three uses on the site. Its lack of a steady direction is the reason we do not correct Australian polls: the leave-one-out test reported here rejects the correction. Its size sets the width of the simulation's national errors (Simulating Australian polling error: where the chances come from). And its pollster-by-pollster detail feeds the ratings (Rating Australian pollsters, and how to read a poll release).
What has changed since this paper was written
Updated 9 November 2026. The rest of the paper is as first published, except where a sentence below says it was corrected.
- When the polls missed (2 November): The miracle that was a sampling problem: Australia's 2019 polling failure.
What is measured, and against what
The quantity that decides the House is each division's count, but what pollsters publish nationally is the first-preference (primary) vote and, usually, a 2PP for Labor against the Coalition. We measure both against the Australian Electoral Commission's national figures, and against the Parliamentary Handbook's before 2004.
Polls are rescaled to 100 without the undecided. Parties are grouped as Labor, the Coalition (Liberal, LNP, Nationals and Country Liberal, whose polls and results are reported together), the Greens, One Nation, and others (independents and every minor party). One Nation is measured only where the poll named it: in 1998, and from 2019 on. Before 2019 pollsters folded it into "others", and so are "others" left unmeasured at those elections, because a poll's "others" then held One Nation and the result's did not.
A party is measured at an election if it is in the result and in at least half of that election's final polls. For the site-wide accuracy figures, which are comparable across the ten countries, a main party is one with at least 5% of the vote, and "others" never count.
Definitions
Let \(v_{p,t}\) be party \(p\)'s share of the formal first-preference vote at election \(t\), and \(y_{j,p,t}\) its share in final poll \(j\), one of the \(n_t\) final polls for that election. The error of a single poll is
The sign convention is the same everywhere on the site: error = poll minus result. A positive error means the polls overstated the party.
The industry error at election \(t\) is the error of the simple average of the final polls,
where \(n_{p,t}\) is the number of final polls reporting the party. Each pollster counts once.
To compare elections we use the average miss per party, the mean absolute industry error over the \(P_t\) parties measured:
Across elections each party's errors are summarised by their mean (direction) and root mean square (size):
where \(s_p^2\) is the variance of the errors about their mean. The identity splits the size of a party's misses into a steady part (bias) and a part that changes from one election to the next.
The gap between the two parties with the most first preferences is scored as
where \(a\) finished first and \(b\) second. A positive gap error means the polls put the leader too far ahead.
For the 2PP the definition is equation (2) applied to Labor's share of the two-party vote: the average of the pollsters' own final 2PP figures minus the AEC's national 2PP.
The evidence: final polls and how they were chosen
A final poll is the last national poll by a given pollster before polling day. Seat polls are excluded: until 28 September 2026 the 2022 figures wrongly included five seat polls (among them North Sydney, Curtin and Fowler), which Wikipedia's 2022 polling page lists under "Polling for individual seats" and our reader had taken for national polls. They made 2022 look like a Labor understatement and are now left out.
The record holds 46 final polls with a 2PP and 39 with primaries:
| Period | Elections | Final polls with primaries | Final polls with a 2PP |
|---|---|---|---|
| 1987–1998 (not used by the model) | 1987, 1990, 1993, 1996, 1998 | 3 | 4 |
| 2001–2004 | 2001, 2004 | 0 | 3 |
| 2007–2025 | seven elections | 36 | 39 |
Table: Final polls on file. The model uses elections from 2000 on.
The older record is thin. Before 2007 we hold one to three polls an election: Newspoll's primaries in 1987 and 1990 (Labor and the Coalition only), Nielsen's full primaries in 1998, and an election-eve 2PP from Newspoll or Nielsen in 1993, 1996, 1998, 2001 and 2004. In 2007 only Newspoll's final primaries survive; the other three 2007 final polls give the 2PP alone. In 2010 we hold two (Newspoll and Roy Morgan). From 2013 there are five to ten an election.
Every final poll is grade C: transcribed from the poll tables on Wikipedia's election articles, with the citation to the original release kept against each poll. We found no pollster archive for these years that is open to read; Roy Morgan's findings site lists no older releases and the Newspoll archive survives only on archive.org, which we do not use. The Australian data, its sources and the law on polls describes the grades.
A worked example: 2025
The 2025 election, on 3 May, had the most final polls. Their primaries, rescaled to 100, and the result were:
| Pollster | Fieldwork end | Labor | Coalition | Greens | One Nation | Others | Poll's miss | 2PP |
|---|---|---|---|---|---|---|---|---|
| RedBridge | 29 Apr | 34.0 | 34.0 | 12.0 | 8.0 | 12.0 | 1.50 | 53.0 |
| DemosAU | 30 Apr | 31.0 | 33.0 | 12.0 | 9.0 | 15.0 | 1.51 | 52.0 |
| Roy Morgan | 2 May | 33.0 | 34.5 | 13.5 | 6.5 | 12.5 | 1.62 | 53.0 |
| YouGov | 1 May | 31.1 | 31.4 | 14.6 | 8.5 | 14.4 | 1.79 | 52.2 |
| Newspoll | 1 May | 33.0 | 34.0 | 13.0 | 8.0 | 12.0 | 1.82 | 52.5 |
| Ipsos | 1 May | 29.5 | 34.7 | 12.6 | 8.4 | 14.7 | 2.14 | 51.0 |
| Resolve Strategic | 28 Apr | 31.0 | 35.0 | 14.0 | 7.0 | 13.0 | 2.22 | 53.0 |
| Essential | 27 Apr | 32.0 | 34.0 | 13.0 | 10.0 | 11.0 | 2.62 | 52.1 |
| Freshwater Strategy | 1 May | 33.0 | 37.0 | 12.0 | 8.0 | 10.0 | 2.70 | 51.5 |
| Spectre Strategy | 28 Apr | 31.0 | 34.0 | 14.5 | 9.5 | 11.0 | 3.02 | 53.0 |
| Average | 31.86 | 34.16 | 13.12 | 8.29 | 12.56 | 52.3 | ||
| Result | 34.56 | 31.82 | 12.20 | 6.40 | 14.98 | 55.2 | ||
| Error | −2.70 | +2.34 | +0.92 | +1.89 | −2.42 | −2.9 |
Table: Final polls and result, House 2025, primary vote and Labor's 2PP, points. "Poll's miss" is the poll's own average absolute error over the five groups.
Labor's average is \(318.6/10 = 31.86\) against 34.56, an error of \(-2.70\) by equation (2). Over the five groups the average miss is \((2.70 + 2.34 + 0.92 + 1.89 + 2.42)/5 \approx 2.06\) points; over the four main parties (others excluded), 1.97. The two largest parties on first preferences were Labor and the Coalition, 2.7 points apart; the polls had the Coalition 2.3 points ahead, a gap error of \(-5.0\) by equation (5), and Ipsos's of \(-7.9\).
Three features of 2025 matter for what follows. The miss was shared: every poll understated Labor, every poll but one overstated the Coalition, and every poll overstated One Nation. The average was no closer than its better members; it missed by 2.06, the individual polls by 1.50 to 3.02. And the 2PP missed by the same amount as the primaries implied: every pollster's 2PP was between 2.2 and 4.2 points too low for Labor.
The record, 1987–2025
The table applies equation (2) to every election with final primaries.
| Party | 1987 | 1990 | 1998 | 2007 | 2010 | 2013 | 2016 | 2019 | 2022 | 2025 |
|---|---|---|---|---|---|---|---|---|---|---|
| Labor | +2.7 | +2.0 | +1.1 | +0.6 | +2.2 | +0.3 | −0.2 | +2.4 | +2.4 | −2.7 |
| Coalition | −3.1 | −4.0 | +3.7 | +0.9 | −2.9 | −1.2 | 0.0 | −2.8 | −0.3 | +2.3 |
| Greens | −0.5 | −0.8 | +3.1 | +1.7 | +0.8 | −0.4 | 0.0 | +0.9 | ||
| One Nation | −0.2 | +1.0 | −0.1 | +1.9 | ||||||
| Others | −4.2 | −0.2 | −2.0 | −2.4 | ||||||
| Final polls | 1 | 1 | 1 | 1 | 2 | 8 | 5 | 5 | 5 | 10 |
| Miss per main party | 2.88 | 2.97 | 1.68 | 0.77 | 2.74 | 1.06 | 0.33 | 1.87 | 0.90 | 1.97 |
Table: Final-poll average minus result, primary vote, points. The last row is equation (3) over the parties with 5% or more of the vote, others excluded. 1987, 1990 and 1998 are shown for the record and are not used by the model.
Four points stand out.
Labor is usually overstated, until it is not. Labor was overstated at eight of the ten elections, and at five of the seven the model uses. The exceptions are 2016, when the polls were almost exact, and 2025, when they had Labor 2.7 points too low. A correction learned from 2007 to 2022 would have lowered Labor's figure in 2025 and made the largest miss larger.
The Coalition's misses are the mirror of Labor's, mostly. It was understated in 2010, 2013 and 2019 and overstated in 2007 and 2025. The two largest parties' errors had opposite signs at seven of the ten elections.
The Greens have leaned high. They were overstated at five of the seven elections the model uses, by 3.1 points in 2010 and 1.7 in 2013. The same lean shows at state level (below).
2010 is the worst year since 2000. Its average miss per main party was 2.74. It rests on two polls, one of them a Roy Morgan poll with fieldwork ending on 8 August, thirteen days before polling day, which had Labor on 44.1% against a result of 38.0. Newspoll's final poll that year missed by 1.38.
Using equation (4) on the seven elections the model uses:
| Party | Elections | Mean error | RMS error | Share of RMS² that is bias |
|---|---|---|---|---|
| Labor | 7 | +0.72 | 1.85 | 15% |
| Coalition | 7 | −0.58 | 1.87 | 10% |
| Greens | 7 | +0.77 | 1.46 | 27% |
| One Nation | 3 | +0.93 | 1.25 | 56% |
| Others | 3 | −1.55 | 1.82 | 72% |
Table: Summary of industry errors, 2007–2025, points. The last column is \(\bar{e}_p^2 / \text{RMS}_p^2\).
For the two largest parties almost all of the error is variation from one election to the next. For One Nation and others a larger share looks like bias, but on three elections each, with the largest misses all in 2025. The RMS column is what the simulation uses as each party's spread.
Shared and pollster-specific error
As in Measuring British polling error in England, 2015–2024, the average squared poll error at an election splits exactly into the square of the industry error and the spread of the polls about their average:
The share of the first term is the share of the miss common to the industry. For the largest Australian misses it is high: 98% for the Coalition in 2019 (every poll between 2.4 and 3.4 points too low), 81% for Labor in 2025, 78% for One Nation in 2025 and 72% for Labor in 2019. No choice of pollster would have avoided these, and averaging more polls does nothing to reduce them. Where the industry was right on average the shared part is small: 2% for Labor and 0% for the Coalition in 2016, 0% for the Greens in 2022.
The 2PP
Most Australian pollsters publish a 2PP, and the pollsters' final 2PP figures have their own record, from 1993:
| Election | Final polls | Poll average | Result | Miss |
|---|---|---|---|---|
| 1993 | 1 | 49.5 | 51.4 | −1.9 |
| 1996 | 1 | 46.5 | 46.4 | +0.1 |
| 1998 | 2 | 51.5 | 51.0 | +0.5 |
| 2001 | 1 | 47.0 | 49.1 | −2.0 |
| 2004 | 2 | 48.0 | 47.3 | +0.7 |
| 2007 | 4 | 53.6 | 52.7 | +0.9 |
| 2010 | 2 | 53.9 | 50.1 | +3.7 |
| 2013 | 8 | 46.8 | 46.5 | +0.3 |
| 2016 | 5 | 49.4 | 49.6 | −0.2 |
| 2019 | 5 | 51.4 | 48.5 | +2.9 |
| 2022 | 5 | 51.6 | 52.1 | −0.5 |
| 2025 | 10 | 52.3 | 55.2 | −2.9 |
Table: Labor's 2PP in the pollsters' final polls against the AEC's national 2PP, points.
Over the nine elections from 2001 the mean is +0.3 and the mean absolute miss 1.6 points; there is no steady direction. 2019 is the miss that is remembered, because every final poll had Labor ahead (51.0 to 52.0) and it lost; in 2025 every final poll was between 2.2 and 4.2 points too low for Labor. Single polls can differ widely: Roy Morgan's 2010 figure was 7.4 points too high for Labor and Essential's 2022 figure 4.1 too low, the latter because Essential's 2PP keeps an undecided share.
The forecast does not use the pollsters' 2PP. It converts primaries to a 2PP itself with the preference flows of the last election (From primaries to seats: preference flows and the 2PP), and that conversion, applied to each election's actual primaries with the previous election's flows, came within about a point of the AEC's figure at every election from 2007 to 2025. So the 2PP record is shown, not corrected.
The gap between the two largest parties
| Election | Top two on primaries | Real gap | Polls' gap error | Worst poll |
|---|---|---|---|---|
| 2007 | Labor, Coalition | 1.3 | −0.3 | −0.3 |
| 2010 | Coalition, Labor | 5.6 | −5.1 | −11.8 |
| 2013 | Coalition, Labor | 12.2 | −1.5 | −4.2 |
| 2016 | Coalition, Labor | 7.3 | +0.1 | +0.9 |
| 2019 | Coalition, Labor | 8.1 | −5.2 | −7.1 |
| 2022 | Coalition, Labor | 3.1 | −2.7 | −4.1 |
| 2025 | Labor, Coalition | 2.7 | −5.0 | −8.0 |
Table: Gap between the two parties with the most first preferences, points, equation (5).
The gap error has been negative at six of seven elections: the polls have usually underestimated the leader's margin on primaries. That is not a steady lean towards one party, because the leader changed: in 2010, 2013, 2019 and 2022 the polls understated the Coalition's lead over Labor, in 2025 Labor's lead over the Coalition. In 2025 the average put the parties the wrong way round on primaries (Coalition 34.2, Labor 31.9).
Pollster by pollster
Each pollster's record is the average miss per main party of its own final polls, at the elections from 2000 it covers:
| Pollster | Elections | Average miss per main party |
|---|---|---|
| Galaxy Research | 2 | 0.83 |
| uComms | 2 | 0.88 |
| ACNielsen / Nielsen | 1 | 1.06 |
| RedBridge | 1 | 1.14 |
| AMR | 1 | 1.17 |
| Newspoll | 7 | 1.22 |
| Resolve Strategic | 2 | 1.77 |
| DemosAU | 1 | 1.89 |
| Ipsos | 4 | 2.05 |
| Essential | 5 | 2.13 |
| Freshwater Strategy | 1 | 2.14 |
| Roy Morgan | 5 | 2.19 |
| YouGov | 2 | 2.30 |
| Spectre Strategy | 1 | 2.79 |
| Lonergan Research | 1 | 3.17 |
Table: Average miss per main party of each pollster's final polls, 2007–2025, points.
Two cautions apply. The records cover different elections: Galaxy and uComms were tested in 2013 and 2016, when the industry did well; every pollster with a 2025 final poll carries that year's shared miss. And the differences are small next to the differences between elections. Newspoll, with the only seven-election record, missed by 0.18 in 2016 and 2.83 in 2019. Roy Morgan's average is lifted by one poll, its 2010 figure (5.30); its other four average 1.41. The ratings (Rating Australian pollsters, and how to read a poll release) pull short records towards a typical final poll for this reason.
Why the record is not used to correct
The obvious use of the record is to subtract each party's average miss from today's polls. Our rule takes off half the average miss, equation (1) of Forecasting Australia's House from imperfect polls: an overview, and keeps the correction only if it would have improved past forecasts, each election predicted with the correction learned from the others:
| Election | Polls' average miss | Corrected | Verdict |
|---|---|---|---|
| 2007 | 0.77 | 1.09 | further out |
| 2010 | 2.74 | 2.59 | closer |
| 2013 | 1.06 | 0.75 | closer |
| 2016 | 0.33 | 0.47 | further out |
| 2019 | 1.38 | 1.45 | further out |
| 2022 | 0.97 | 1.02 | further out |
| 2025 | 2.06 | 2.06 | further out |
| Mean | 1.33 | 1.35 |
Table: Leave-one-out test of the λ = ½ correction, average miss per party over the parties the model tracks (Labor, Coalition, Greens, One Nation, others), points.
Correcting was closer at two elections and further out at five, and worse on average. The correction learned from all seven would lower today's Labor figure by 0.36 points and raise the Coalition's by 0.29; the test says not to apply it. The record is used instead to set the size of the uncertainty (Simulating Australian polling error: where the chances come from).
Until 28 September 2026 the test also used 1987, 1990 and 1998. Over those ten elections the correction passed (1.62 against 1.71, closer at six of ten), almost entirely because of the two single Newspolls of 1987 and 1990. The owner's rule that elections before 2000 shape nothing removed them, and the verdict reversed.
State elections as a check
State elections are polled by many of the same firms. For the two states whose elections the site forecasts (Victoria on 28 November 2026 and New South Wales on 13 March 2027), the database holds the final state polls of eight elections, from 1 to 9 polls each:
| Election | Labor | Coalition | Greens | Labor 2PP |
|---|---|---|---|---|
| Victoria 2010 | −3.3 | +0.2 | +3.8 | +0.5 |
| NSW 2011 | −2.8 | +0.7 | +1.7 | −0.8 |
| Victoria 2014 | −1.0 | −2.0 | +3.0 | +0.2 |
| NSW 2015 | −1.1 | +0.1 | +1.1 | −0.8 |
| Victoria 2018 | −3.0 | +2.0 | +0.4 | −3.2 |
| NSW 2019 | +2.5 | −2.2 | +0.1 | +2.0 |
| Victoria 2022 | +0.4 | +0.1 | 0.0 | −0.9 |
| NSW 2023 | −0.2 | +0.6 | +0.8 | −0.6 |
Table: State final-poll average minus result, points.
The Greens were overstated at all eight (by only 0.03 points in Victoria in 2022), which supports the federal lean; Labor was understated at six of eight, the opposite of the federal lean. These figures feed the state forecasts, not the federal one, and are shown here as a check between federal elections.
Limits
Seven elections. Every summary rests on seven numbers per party, three for One Nation and others. A single election like 2025 moves the means a long way.
Few polls in early years. 2007 has one poll with primaries and 2010 two. An error "of the industry" in those years is one or two pollsters' error.
Secondary sources. Every final poll is transcribed from Wikipedia's tables. A transcription error would pass straight into the record. The 2022 seat-poll mistake, found and fixed, shows that this happens.
One Nation and others. They have been measured only since 2019, and the size of their past misses says little about polling a party at 26%.
Final polls only. These errors say nothing directly about polls taken long before an election; that is Polling day: how far Australian polls move.
References
Australian Electoral Commission, national results and two-party-preferred, 2004–2025 (CC BY 4.0); Parliament of Australia, Parliamentary Handbook, results before 2004.
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.
How to cite
Polling Forecast (2026). How wrong have Australian polls been? Primaries, the 2PP and the λ test. Retrieved 21 November 2026 from https://www.pollingforecast.com/research/australia-polling-error. Live figures change as new polls arrive; the date above is when the text was last revised.



