Rating Australian pollsters, and how to read a poll release

Summary

We rate each Australian pollster out of 100: 50 points for the accuracy of its final House polls since 2000, 20 for its error on the gap between the two parties that finished first and second, and 30 for how much of its working it publishes. Accuracy is scored by rank against all 36 Australian final polls from 2007 to 2025, and a pollster with a short record is pulled towards the typical one, so a single good election cannot earn a top grade. On 15 October 2026 the best-rated pollsters polling now are RedBridge (65, provisional), Newspoll (62) and DemosAU (62, provisional); Galaxy Research, which no longer polls under that name, has the best record (75). The rating is shown for readers and does not weight the polls: the average gives every pollster's latest poll the same weight. This paper also sets out each pollster's lean, past and present, and how to read an Australian poll release, where the undecided share, the 2PP method and the state breakdown all need care.

What the rating is for

The rating answers a narrow question: on the evidence we can check, how far a reader should trust a pollster's published figure, and how easily they can see how it was made. It does not say which pollster is right today, and it does not feed the forecast. The Australian average is the plain mean of each pollster's latest national poll within 21 days of the newest, one poll per pollster (Forecasting Australia's House from imperfect polls: an overview). Roy Morgan, which publishes weekly, counts no more than Essential, which publishes fortnightly.

We do not weight by the rating for the same reason we do not correct Australian polls: the evidence is seven elections, most pollsters have one or two of them, and the records move from one election to the next (How wrong have Australian polls been? Primaries, the 2PP and the λ test).

The evidence: 36 final polls

The accuracy parts of the rating use every final poll with primaries at the House elections since 2000: one in 2007, two in 2010, eight in 2013, five each in 2016, 2019 and 2022, and ten in 2025. Fifteen pollsters have at least one. The same 36 polls set the country's yardstick.

For pollster \(j\) at election \(t\), write \(e_{j,t,p} = y_{j,t,p} - v_{t,p}\) for its share for party \(p\) minus the result. The miss is the mean absolute error over the main parties it reported, those with at least 5% of the vote, never "others":

The lead error is the error on the gap between the two parties that actually finished first and second on first preferences:

In Australia the pair is always Labor and the Coalition, in the order of the result: the Coalition first from 2010 to 2022, Labor first in 2007 and 2025.

Scoring

The country's yardstick

Each country is scored against its own final polls, because a typical miss differs a good deal between countries. The 36 Australian misses, sorted, give the yardstick:

Table: The Australian yardstick, points, from the 36 final polls of 2007–2025.

Shrinkage

Each pollster's record is pulled towards the typical final poll as if it had two more elections at that level:

where \(n_j\) is the number of the pollster's scored final polls, \(\bar{m}_j\) and \(\bar{\ell}_j\) its averages of \(m_{j,t}\) and \(|\ell_{j,t}|\), and \(m_{\text{typ}} = 1.58\) and \(\ell_{\text{typ}} = 3.08\) the medians. With one election the typical poll carries two-thirds of the weight; with seven, two-ninths.

Points by rank

Let \(R(x)\) be the share of the country's final polls that missed by less than \(x\), interpolated between neighbouring polls. Then

A pollster whose rated miss equals the median scores \(50 \times 0.5^{0.75} \approx 30\) accuracy points; one in the best tenth about 46. The power of 0.75 gives more credit for beating most rivals than a straight line would. Because the yardstick is Australia's own, grades show how a pollster compares with its Australian rivals and cannot be compared across countries.

Openness

The 30 openness points reward publishing enough that a figure can be checked:

Table: Australian openness items.

Each item is Yes, Partial or No, worth the full points, a third, or nothing. An item not yet checked counts neither for nor against: the points are scaled to 30 over the items checked,

where \(C_j\) is the set of items checked for pollster \(j\). A pollster whose publications have not been checked at all is graded on its record alone, \((A_j + L_j) \times 100/70\), and marked "record only".

Grades

The score is \(\operatorname{round}(A_j + L_j + O_j)\), graded A (75 and above), B (60–74), C (45–59), D (30–44) and E (below 30). A rating is provisional when the pollster has fewer than two scored final polls.

A worked example: Newspoll

Newspoll has the only final poll at all seven elections. Its misses per main party were 0.77 (2007), 1.38 (2010), 0.39 (2013), 0.18 (2016), 2.83 (2019), 1.45 (2022) and 1.54 (2025), an average of 1.22. Its lead errors were −0.3, +1.6, +0.8, −0.3, −7.1, −4.1 and −3.7, an average absolute error of 2.56. By equation (3),

A rated miss of 1.30 beats about 64% of Australian final polls, which by equation (4) is worth 36 accuracy points; the rated lead error is worth 13. Newspoll publishes a long methodology statement for each poll with its sample, dates, undecided share and 2PP method, and Pyxis, which runs it, belongs to the Australian Polling Council; its results appear in a paywalled newspaper with no full tables, and no raw figures. Its openness is 14 of 30, and its score \(36 + 13 + 14 = 62\), a B.

The current table

Table: Pollsters with a final-poll record and a national poll in the last year, 15 October 2026.

Three newer pollsters (Fox & Hedgehog, Agenda C Synesis and Wolf & Smith) have published national polls since 2025 but have no final poll yet, so they are not graded. Six pollsters have a record but no poll in the last year: Galaxy Research (75, A, record only, now part of YouGov), uComms (73, B), ACNielsen (66, B, provisional), AMR (64, B, provisional), Ipsos (50, C) and Lonergan (48, C, provisional).

The spread matters more than the letters. Roy Morgan's accuracy points are pulled down by one poll, its 2010 figure, which missed by 5.30 points per main party; its other four average 1.41. Essential's openness is among the highest, but its record includes 2022, when its primaries, published with an undecided share, missed by 3.55. The provisional pollsters' points sit close to the typical poll's, because one election carries a third of the weight.

Each pollster's lean

At past elections

The mean error of each pollster's final polls, by party:

Table: Mean error of final polls, main parties, points. Positive: the pollster had the party too high. One Nation is scored only in 2025, the one election since 2000 at which it had 5% of the vote.

Every pollster with a 2025 final poll overstated One Nation that year, so every One Nation figure here is positive.

Now

A pollster's lean now is its latest poll within the 21-day window minus the average:

Table: Latest poll minus the average, points, 15 October 2026.

Because the average is built from the same polls, the leans roughly cancel. A lean is not an error: nobody knows yet who is right. Roy Morgan has One Nation lowest and others highest; RedBridge has One Nation highest and the Greens lowest. Readers can weigh a lean against the pollster's past errors: Roy Morgan, for example, understated the Coalition at its past final polls by 1.6 points on average and sits close to the average on the Coalition now.

How to read an Australian poll release

Primary vote and undecided. Most pollsters ask a "leaner" question of those who say they are undecided, then exclude the rest: Newspoll excluded 5% in November 2025, RedBridge excludes 6 to 9%. Roy Morgan publishes its "can't say" share. Essential publishes primaries with the undecided left in. We rescale every poll's primaries to 100 without the undecided, so polls that exclude different shares can be averaged.

The 2PP. A pollster's 2PP is worked out either with the preference flows of the last election, or from respondents' stated preferences. Newspoll uses previous-election flows; Roy Morgan headlines a respondent-allocated figure and also publishes one on 2025 flows; RedBridge has published both since March 2026; Essential's "2PP+" keeps an undecided share. The two methods can differ by several points when a party's voters change how they rank the others, as One Nation's may have done. We work out our own 2PP from the primaries and show the pollsters' beside it (From primaries to seats: preference flows and the 2PP).

Labor against One Nation. Since One Nation passed the Coalition in the polls, several pollsters publish a Labor v One Nation 2PP as well. These are the only published evidence on how Coalition voters rank One Nation against Labor.

State breakdowns. Newspoll and YouGov publish state figures from their national samples, which are small in each state. The forecast uses half of their swing beyond the national swing, and only where two pollsters agree to publish (The division model and its tests).

The Australian Polling Council. Members commit to publishing a long methodology statement for each poll, with the sample, the weighting, the question wording and the 2PP method. Newspoll (Pyxis), Essential, DemosAU, YouGov, Fox & Hedgehog, Ipsos, uComms and Lonergan are members; Roy Morgan, Resolve, RedBridge, Freshwater, Spectre, Agenda C and Wolf & Smith are not.

Paywalls. Newspoll's and Resolve's figures appear in paywalled newspapers, and DemosAU's site answers with a captcha, which we do not bypass. Their polls are graded C, from Wikipedia's tables, unless the release can be read (The Australian data, its sources and the law on polls).

What the rating leaves out

The rating leaves out how often a pollster polls, who commissions it, which way it leans today, and its mode. It also leaves out seat polls and MRP models, which several Australian pollsters publish but which are not national voting intention.

Limitations

Short records. Seven pollsters have one scored final poll, and Newspoll alone has more than five. The shrinkage limits what a short record can say, but it does not correct for which elections a pollster happened to cover.

Which elections. Galaxy and uComms were tested in 2013 and 2016, when the industry did well; every 2025 pollster carries that year's shared miss of Labor.

The sources. Every final poll is grade C, transcribed from Wikipedia. A transcription error would pass straight into a rating.

Methods change. Newspoll has been run by three different firms (Newspoll Market Research, then YouGov and Galaxy from 2015, Pyxis since 2023); Galaxy merged into YouGov. A record built on 2007–2025 may describe a pollster that no longer works that way.

Openness is what we could check. Several items for several pollsters are unchecked and count for nothing either way.

References

Australian Polling Council, member list and disclosure code.

Efron, B. and Morris, C. (1975). Data analysis using Stein's estimator and its generalizations. Journal of the American Statistical Association, 70(350), 311–319.

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