Victoria88-26: the Victorian application
How the 98th Percentile model is applied to the 2026 Victorian state election. The data, the calibration, the assumptions, and what is published when.
Version 1.0, 27 October 2026. This page is the companion to the model. That page describes the method, which does not change between elections. This page records every choice specific to the 2026 Victorian state election, published as Victoria88-26, and it is the page that changes as the campaign proceeds. It is versioned, and changes are recorded in the change log.
The election
Victoria elects its Legislative Assembly on 28 November 2026: 88 single-member electorates under full preferential voting, where every square must be numbered and ballots essentially never exhaust. The election is fought on the boundaries drawn at the 2021 redivision, the same boundaries as 2022, so no notional recalculation of past results is needed this cycle. The roll closes at 8 pm on 3 November, when the writ issues, and nominations close six days later on 9 November under the Electoral Amendment Act 2026, with Legislative Council group registration closing at noon that day. Early voting cannot begin before 18 November. The candidate list and the registered how-to-vote cards become knowable once nominations close.
The same Act requires a full distribution of preferences in all 88 Legislative Assembly districts, so for the first time the complete preference count will be published for every seat rather than a two-candidate-preferred count in most.
The Legislative Council is not forecast. Upper houses are outside the model’s scope everywhere, and Victoria adds a reason of its own: the Electoral Amendment (Miscellaneous Matters) Act 2026, passed on 11 August 2026, removed group voting tickets from Legislative Council elections, and voters must now number at least five boxes above or below the line. That system has no Victorian precedent to calibrate against, and because a ballot may stop at five preferences, Council preferences can exhaust.
The central fact of this election is One Nation. The party polls at several times its best previous Victorian result, a rise with no precedent in the state’s history, and the model’s handling of it is the single largest determinant of this forecast’s quality. Most of this page is about that.
The polls
Six pollsters are publishing Victorian state voting intention this cycle: Redbridge and Accent Research, Resolve Strategic, Freshwater Strategy, DemosAU, Roy Morgan and Newspoll conducted by Pyxis. Each poll is captured from the pollster’s own release, with Australian Polling Council disclosure statements used for methodology detail where available and the consolidated Wikipedia series used to identify releases. Every capture is timestamped and nothing is overwritten, so a data revision can always be distinguished from a model change.
Poll series are held for five cycles, 2010 through 2026. The four completed cycles are what anchor the industry offset; without them the offset would rest on its prior alone.
The structural break. Victorian pollsters did not prompt for One Nation in their state voting-intention question until early 2026. A party absent from the prompted list is systematically understated, so the change is a change in the measuring instrument, not a movement in opinion, and it affected every pollster at roughly the same time. The model treats 1 February 2026 as a hard discontinuity: no poll before that date informs the One Nation series. The date is the conservative bound, set from the first clean single-wave poll that prompted for the party. One earlier poll blending January and February fieldwork is excluded entirely rather than used to pull the break earlier.
The consequence is that the One Nation series is short. Every other party’s latent series draws on five cycles of polling; One Nation’s draws on the months since February 2026. One electoral observation falls after the break, the Nepean by-election of 2 May 2026, the only occasion on which a Victorian electorate has actually voted with One Nation polling at its current level. It does not enter the model: the latent series is anchored to general elections, and the electorate layer is built from federal, Census and South Australian data. It stands as an outside check, read with the caution that by-elections overstate protest swings.
Anchoring. The latent series is anchored to the Victorian Electoral Commission’s results for 2010, 2014, 2018 and 2022. Narracan’s 2022 election failed after the death of a candidate and was decided at a supplementary election in January 2023; that result is used. Four anchors is few, which is why the prior on the offset’s between-election variance carries real weight in this application, as the model explains.
The One Nation layer
One Nation’s statewide level comes from the aggregator, measured directly by the post-break polls. What the polls cannot supply is the geography: where in Victoria that vote sits. The party’s own Victorian history is useless for this, because the relationship between place and One Nation vote was never observed at anything like the current level of support. The layer is therefore built from two adjacent sources and disciplined by a third.
The federal map. The starting point is One Nation’s vote at the 2025 federal election, moved from polling places onto the 88 state electorates using the Australian Electoral Commission’s count of votes by Census small area.
The education gradient. Each electorate’s federal result is compared with the share of its adults who completed Year 12. Where the federal result is unusually high or low for the electorate’s education profile, the model pulls it part of the way back towards what the profile predicts.
The South Australian calibration. How far to pull is not guessed. It is estimated from the 2026 South Australian election, the one place the same surge has already run its course on comparable geography, and transferred to Victoria with its uncertainty deliberately inflated.
South Australia earns that role because it is a complete, resolved observation of exactly the problem Victoria now poses, and because it contains a warning. Queensland 1998 is the older precedent, but it is nearly three decades old, pre-dates the current party system and offers no comparable Census geography; South Australia is contemporary, measured by the same polling industry, on current boundaries. Before that election, the available federal mapping pointed to regional and rural strength, with inner Adelaide weak and the country strong (Green 2026b). The state result landed differently: One Nation polled above 30 per cent in a band of outer-suburban, low-income, historically safe Labor divisions in Adelaide’s north, territory the federal gradient had not flagged (Green 2026a). Our own analysis of the Legislative Council first preferences by division, where every group contested all 47 divisions and candidate-selection bias therefore vanishes, found the gradient was socioeconomic rather than partisan: One Nation ran strongest where Greens support was weakest and the informal vote highest, an education and affluence axis that cuts across the Labor-Liberal divide and collects whichever party holds low-income territory, Labor in the outer suburbs and the Coalition in the country.
Two further South Australian facts discipline the Victorian layer. One Nation’s division-level dispersion was unremarkable, more spread than Labor’s and less than the Liberals’, so the party is not geographically flat and a flat allocation would be wrong. And its vote had a ceiling: above 30 per cent in 13 divisions, above 40 in none, which under full preferential counting makes preferences decisive and belongs to stage three rather than to geography.
The back-test. An earlier, simpler version of this layer was tested on South Australia directly: given each division’s federal result and the statewide total, it reproduced the division’s One Nation vote with a mean absolute error of 1.73 percentage points across the 46 divisions tested, against a standard deviation across divisions of 8.16 points. The current version builds on the one tested and cannot itself be tested on South Australia, because South Australia now informs its priors. Its first genuine test is the Victorian result.
Who loses the votes. The model must also decide which parties shrink as One Nation grows, because that determines the exclusion order in every seat and therefore which two candidates reach the final count. Two variants are run. The primary variant takes One Nation’s above-average gains disproportionately from Labor, weighted by how much vote each party has available to lose. Its coefficient comes from our division-level analysis of the South Australian change from 2022 to 2026: where One Nation gained most, Labor lost most, at roughly half a point of Labor vote per point of above-average One Nation gain, and in the strongest divisions the arithmetic is decisive on its own, since the combined conservative vote there was too small to have supplied what One Nation won. The alternative variant takes the gains from every party in proportion to its vote. Neither is validated for Victoria; the difference between them is reported with every release. South Australia redistributes its boundaries between every election, so the 2022-to-2026 division pairs behind the coefficient are inexact; the full attribution analysis, with its caveats, is being prepared as its own article.
If the transfer holds, the Victorian exposure is the northern and western growth corridors, safe Labor, low-income, outer suburban, with regional Victoria the second band. If it does not hold, the error will be geographic rather than statewide: the level would stand while the map beneath it moved.
Independents
Until nominations close on 9 November, the independent field is drawn from 2022’s: whether each sitting independent recontests, at what share, plus a draw for new entrants. This matters most in the handful of electorates an independent already holds. The drawn field is replaced by the real one when nominations are declared.
Preferences and seats
Stage three is designed and not yet published. Transfer proportions will be drawn from priors estimated from the Victorian Electoral Commission’s distributions of preferences for 2014 through 2022, updated with the how-to-vote cards registered after nominations close.
The South Australian election shows why this stage cannot be skipped and why it must wait for the cards. One Nation’s 22.9 per cent of the statewide lower-house vote returned four seats of 47, fewer than the Liberals won on 18.9, because a seat is won by the candidate holding a majority after the distribution of preferences, so a party whose vote peaks in the thirties must find the balance from other parties’ preferences, and One Nation was widely preferenced last. Kevin Bonham’s analysis of the flows puts Greens preferences flowing overwhelmingly to Labor in both Labor-Liberal and Labor-One Nation contests (Bonham 2026).
Victoria already shows the same sensitivity. An MRP conducted by RedBridge and Accent Research in June 2026 projected 39 Coalition seats, 35 Labor, 12 One Nation and 2 Greens. Ben Raue, applying his own preference-flow estimates to the same primary votes of 27 per cent One Nation, 26 each for Labor and the Coalition and 13 Greens, projected 31 Coalition, 30 Labor and 27 One Nation (Raue 2026). The two differ by 15 seats for One Nation on identical first preferences. A first-preference lead and a seat are different things, and the gap between them is set by preference behaviour that is unknowable before the candidate field and the cards exist. That is why this site publishes first-preference estimates now and win probabilities only after 9 November.
Assumptions of this application
These are graded on the scale defined in the model’s assumptions, and they come on top of the method assumptions there.
| # | Assumption | Grade |
|---|---|---|
| V1 | Polls measure One Nation about as accurately in Victoria as they did in South Australia in March 2026 | Reasoned, on one election |
| V2 | One Nation’s geographic distribution in Victoria resembles its 2025 federal distribution, adjusted towards the Census education gradient observed in South Australia | Reasoned, on one election |
| V3 | One Nation’s above-average gains come disproportionately from Labor, on South Australian coefficients | Assumed; an alternative is run and reported |
| V4 | One Nation contests all 88 electorates | Assumed; false until its candidate list is confirmed |
| V5 | The independent field resembles 2022’s until nominations close | Assumed; replaced when nominations are known |
V1 rests on the South Australian pollsters’ final-week record, which was close to the result. V2 is the education-gradient construction above, and carries the inflated transfer uncertainty. V3 is the vote-source choice, with the proportional alternative reported alongside it. V4 overstates One Nation’s effective vote until its slate is confirmed, in proportion to however many electorates it leaves uncontested; nominations close on 9 November. V5 expires the same day.
V1 to V3 rest on a single adjacent election, and no amount of Victorian polling before election day can retire them. With V4, the slate assumption that expires at nominations, they are this forecast’s principal exposure.
Data sources
- Opinion polls. Pollsters’ own releases; Australian Polling Council disclosure statements; the consolidated Wikipedia series as an index of releases. Five cycles, 2010 to 2026.
- Election results. Victorian Electoral Commission first preferences, two-candidate-preferred counts, distributions of preferences and polling-place results, 2010 to 2022, including the Narracan supplementary election.
- Federal results. Australian Electoral Commission 2025 House of Representatives results, with the votes-by-SA1 correspondence used to move them onto state boundaries.
- South Australian calibration data. Electoral Commission SA results for 2022 and 2026, including the Legislative Council first preferences by division.
- Census and geography. ABS 2021 Census at small-area level; ABS state electoral division boundaries. The 2022 and 2025 boundary files are geometrically identical for Victoria. The 2021 redivision abolished nine electorates, created nine, and left only eight of the surviving 79 untouched, so no comparison on this site crosses that boundary without redistribution.
- Known limitations of the data. Pre-2026 Victorian polls carry no One Nation figure. Some older poll records have month-level fieldwork precision. A small number of Roy Morgan records before 2014 have their survey mode inferred rather than confirmed, and are marked as such.
The current release
Published now: statewide first-preference estimates for every party, the same estimates for each of the 88 electorates, and for each electorate the probability of holding the highest first-preference vote, all with plausible ranges as defined on the model page, with the full table available as a CSV. The run identifier, data cut-off and poll count for the current release are printed with every figure. [RUN DETAILS: to supply with the next release.]
Not published until after nominations close and the how-to-vote cards are registered: win probabilities and seat totals, for the reasons in Preferences and seats.
Where a count of leading electorates is published, the electorates sitting close enough to even that simulation noise could move them are named with it, per the model’s known limitations.
Change log
Version 1.0, 27 October 2026. First publication. Covers the aggregator and electorate layers as applied to Victoria; the preference layer is designed and unpublished.
References
Use of AI
I used large language models, Anthropic’s Claude Opus 5.5 and Claude Fable 5, to test this work for narrative flow, logical consistency, bias and errors. They also assisted with drafting and with the site’s code. Responsibility for the content is mine alone.
John Ward (The 98th Percentile)
Copyright
Citation
@online{ward_(the_98th_percentile)2026,
author = {Ward (The 98th Percentile), John},
publisher = {The 98th Percentile},
title = {Victoria88-26: The {Victorian} Application},
date = {2026-10-27},
url = {https://the98thpercentile.pages.dev/methodology-victoria.html},
langid = {en-AU}
}