What we don't know yet: limitations of the Australian forecast
Abstract. The Australian forecast rests on seven elections of final polls, two of them with one or two polls each. It has no precedent for a party at One Nation's level: how its vote spreads across divisions, how Coalition voters rank it against Labor, and whether it keeps its gain to polling day cannot be tested on past elections. The simulation holds preference flows fixed; the division model keeps the 2025 candidates and boundaries; state polls are few; the Senate count is simplified; most polls are second-hand. For each we say what evidence would change our mind.
Summary
The Australian forecast makes a narrow claim: the polls, read as published, carried through a replay of the preference count in all 150 divisions that reproduces 149 of the 2025 winners, with the uncertainty of past polling misses around them. That leaves less to go wrong in the reading of the polls, and more in the step from votes to seats, at a time when that step has become unusually hard. The forecast rests on seven elections of final polls, two of them with one or two polls each. It has no precedent for a party at One Nation's level: how its vote spreads across divisions, how Coalition voters rank it against Labor, and whether it keeps its gain to polling day are the three largest uncertainties, and none can be tested on past elections. The simulation holds preference flows fixed. The division model keeps the 2025 field of candidates and the 2025 boundaries. The state polls behind the state swing are few. The Senate count is a group-level simplification. And most polls reach us second-hand.
This paper sets out each weakness, how much we think it matters, and what evidence would change our mind. The British counterpart is What we don't know yet: limitations and open questions; the method is set out in Forecasting Australia's House from imperfect polls: an overview.
Seven elections
Everything the forecast knows about the size of Australian polling error comes from the final polls at seven House elections, 2007 to 2025: 36 polls with primary votes. The early years are thin. In 2007 one poll's primaries survive (Newspoll's), in 2010 two (Newspoll's and Roy Morgan's); the "industry error" in those years is one or two pollsters' error. One Nation and others have been measured at three elections.
Two consequences follow. First, single elections move the summaries a long way. Without 2025, Labor's mean error would be +1.28 and its root-mean-square error 1.67, against +0.72 and 1.85 with it; a correction learned from 2007–2022 would have lowered Labor's figure in 2025 and made the year's largest miss larger. Second, the leave-one-out test that rejects a correction (How wrong have Australian polls been? Primaries, the 2PP and the λ test) is close: 1.35 points per party corrected against 1.33 uncorrected over seven elections. On the six elections to 2022 alone it was closer still (1.203 against 1.208), short of the 0.01-point margin the rule requires. The verdict could change with one more election.
There is also a structural limit: all polls at an election share the same electorate, campaign and late movement, so 36 polls carry about seven elections' worth of information, not 36.
What would change our mind. Each new House election is added to the record and the test rerun automatically; if a correction began to beat the plain average by the required margin, the code would apply it without any change to the method.
One Nation
One Nation went from 6.4% of first preferences in 2025 to about 26% in today's average. Nothing like it is in the record the forecast is tested on. Three parts of the forecast depend on how One Nation's vote behaves, and none can be checked.
Where its vote lands. The division model moves every candidate by half a uniform and half a proportional swing, with the uniform half never below the proportional figure (The division model and its tests). For a rising party that means half and half where One Nation was below its 2025 national average, and fully proportional where it was above. The election-to-election tests from 2004 to 2022 could not tell the swing rules apart (783 to 784 winners right whatever the share taken as uniform), because no party's vote moved far in those years. If One Nation's new vote is spread more evenly than the model assumes, it would win more of the divisions where it was weak and fewer where it was strong.
How preferences flow. The weights between groups come from the 2025 count, when One Nation had 6.4%. The weight that decides most of today's contests, how Coalition voters split between One Nation and Labor once the Coalition is out, rests on few 2025 divisions, and the same flow fitted on each election alone has swung between 42% and 82% to One Nation since 2004 (From primaries to seats: preference flows and the 2PP). In Dickson, One Nation wins if more than about two-thirds of the Coalition's votes reach it; the model sends 77%. The pollsters' own Labor v One Nation 2PP figures suggest a weaker flow than ours.
Whether it lasts. The election-day forecast pulls every party 69% of the way back to its 2025 result, because that came closer than the polls at the House elections from 2010 to 2025 (Polling day: how far Australian polls move). For One Nation it gives 12.6% on polling day and 4 seats. The record of populist-right parties abroad, and One Nation's own faded rise before 2019, point in opposite directions.
How well it is polled. Its record is three elections: overstated by 1.0 in 2019 and 1.9 in 2025, almost exact in 2022. Polling a party at 26% may be different from polling one at 5%.
What would change our mind. State elections with One Nation standing widely (Victoria in November 2026, New South Wales in March 2027) will show how its vote spreads and how preferences flow to and from it at this level; the distributions of preferences from those counts can be fitted the same way. More pollsters publishing Labor v One Nation and Coalition v One Nation 2PP figures would allow a check of the flows.
Preferences held fixed in the simulation
The simulation draws national, state, candidate and classic-seat 2PP errors, but holds the preference weights and 2PP flow rates at their 2025 values in every run (Simulating Australian polling error: where the chances come from). A shift in flows would move many divisions in the same direction at once, which is exactly the kind of error that widens seat ranges most. The ranges and the combination chances are therefore narrower than they would be if flow uncertainty were included.
What would change our mind. A version of the simulation that draws the key weights from their spread across past elections, tested on the election-to-election replays, would show how much wider the ranges should be.
People and boundaries
The 2025 field. The forecast moves parties, not people. Members who have changed party since 2025 (Barnaby Joyce to One Nation; Zali Steggall and Allegra Spender to Community Strong Australia) keep their 2025 votes with the party they stood for then. New independents, retirements and the personal vote of new members are not modelled. One Nation's win in the Farrer by-election in May 2026 is shown as evidence, not used.
Boundaries. The 2025 divisions are used. Redistributions in South Australia, Tasmania and the ACT, and in Queensland, are under way; each state keeps its number of seats, but notional results for the new boundaries are not in the data (The Australian data, its sources and the law on polls).
Local swing. The candidate error in the simulation is random and independent from division to division. Real swings cluster: outer suburbs, regional Queensland and inner cities tend to move together.
What would change our mind. The AEC's notional results for the new boundaries, as soon as they are published; candidate data for the next election as nominations close.
State polls
The state adjustment adds half of a state's swing beyond the national swing, where at least two pollsters have published state figures within 90 days of that state's newest state poll. On 15 October 2026 five states rest on two pollsters each (Newspoll and YouGov, both from 21 September), and Tasmania on three polls from 1 February to 3 March 2026, seven months old, because the 90-day window runs back from each state's own newest poll rather than from today. The ACT and the Northern Territory have none.
The evidence for the state swing is mixed. With the true 2025 state results it helped (134 winners against 133); with the state polls published before 2025 it did not (127 against 128) (The division model and its tests). It is kept as a halved adjustment.
What would change our mind. A test on more elections. If state polls continued not to help, the adjustment would be dropped; if they helped, the halving could be relaxed.
The Senate
The Senate count is by group, with one set of weights and a fixed 30% exhaustion. It gets 37 of 40 seats right on the 2025 votes, but its misses are last seats, which is where Senate contests are decided (The half-Senate: a simplified quota count). Personal votes (Jacqui Lambie, David Pocock, Lidia Thorpe) are moved with independents or others. In the Senate results the joint Liberal and National tickets carry a code that the grouping table currently assigns to "others", so as of 15 October 2026 they move with the others' change, not the Coalition's; mapped to the Coalition, the same count gives the Coalition two fewer seats.
What would change our mind. A candidate-level count using the AEC's ballot-paper data for 2025, tested against the real result.
The election date and polling day
The forecast assumes the latest usual date, 20 May 2028. An earlier election would change the polling-day forecast sharply: the weight on today's polls is 0.31 a year or more out and 0.62 at 60 days. The pull-back is fitted on six elections, the longest distances on fewer polls, and applies the same curve to every party.
What would change our mind. Each election adds one more set of distances. A party-specific curve, or one that treats new surges differently, would need more cases than the Australian record has; evidence from other countries could be used if it passed the same leave-one-out test.
Second-hand polls
Of the 172 national polls since the 2025 election, 116 are grade C, from Wikipedia's tables, because Newspoll and Resolve publish in paywalled newspapers and DemosAU's site blocks automated reading. Every final poll at past elections is grade C. Where releases could be compared, Wikipedia's rows agree with them, but a transcription error would pass into the average, the record and the ratings. The 2022 record contained five seat polls read as national polls until September 2026.
What would change our mind. Pollsters publishing their own series in a form that can be read, which would let us grade more polls A and collect them automatically.
Pollster ratings
The ratings rank pollsters against Australia's own 36 final polls and pull short records towards the typical poll (Rating Australian pollsters, and how to read a poll release). Seven of the fifteen rated pollsters have one election; Newspoll alone has more than five. The records cover different elections, and every pollster with a 2025 final poll carries that year's shared miss. Grades show how a pollster compares with its Australian rivals and cannot be compared across countries.
Summary
| Limitation | How much it matters | What would address it |
|---|---|---|
| Seven elections, thin early years | The correction verdict is close and could change | More elections; the test reruns itself |
| One Nation's spread across divisions | Could move dozens of seats | State elections with One Nation standing widely |
| Coalition voters' preferences between One Nation and Labor | Decides most three-way contests | Distributions of preferences from new counts; pollsters' Labor v One Nation 2PP |
| Pull-back for One Nation | Drives the polling-day forecast | Cases of large mid-term surges, tested out of sample |
| Preferences fixed in the simulation | Ranges narrower than they should be | Draw flows from their past spread |
| The 2025 field of candidates | New independents and party changes missed | Candidate data at nomination |
| Boundaries | Redistributions not yet in | AEC notional results |
| State polls | Few, from two pollsters; did not help in 2025 | Test on more elections |
| Senate simplification and the Coalition's ticket code | Last seats uncertain; Coalition seats affected | Candidate-level count; the grouping fixed |
| Second-hand polls | Transcription errors pass through | Pollsters' own series |
Table: Known limitations of the Australian forecast.
The next House election will be the first real test of the method as it stands, and we will publish how it went.
References
Jennings, W. and Wlezien, C. (2018). Election polling errors across time and space. Nature Human Behaviour, 2(4), 276–283.
Figures are as of 15 October 2026. The site's pages update with each new poll.
How to cite
Polling Forecast (2026). What we don't know yet: limitations of the Australian forecast. Retrieved 17 October 2026 from https://www.pollingforecast.com/research/australia-limitations. Live figures change as new polls arrive; the date above is when the text was last revised.