The 98th Percentile election forecasting model
A method for Australian lower-house elections. What it estimates, from what data, by what method, and how far each part can be trusted.
Version 1.0, 27 October 2026. This page describes the method in use. It is versioned, and changes are recorded in the change log with their consequences for figures published earlier.
In plain language
This page describes a general method for forecasting Australian lower-house elections. Its first application is the 2026 Victorian state election, published as Victoria88-26. Everything specific to that application - its data sources, its calibration and its particular weaknesses - has its own page, Victoria88-26: the Victorian application, and on this page individual states appear only as examples.
The model works in three stages. The first pools every published voting-intention poll for the jurisdiction into a single daily estimate of each party’s vote across it, adjusting for the fact that different pollsters consistently read parties higher or lower than one another, and for the fact that the polling industry as a whole has sometimes been wrong in the same direction at once. The second stage distributes each party’s jurisdiction-wide vote across the individual electorates, using how each area voted at a recent election and what the Census says about it. The third stage simulates the distribution of preferences to work out who wins each seat; its outputs are published only once the real candidate field and preference recommendations are known.
Estimates are published as probabilities rather than predictions. Every published figure carries a plausible range, which is this site’s plain-language name for a 90 per cent credible interval: the model puts a nine-in-ten chance on the result falling inside it. That is a Bayesian statement, and What the plausible range means sets out what it does and does not entitle a reader to conclude.
Two things are deliberately absent everywhere. The model never uses a two-party-preferred figure, because a two-party summary describes a two-party contest and several Australian jurisdictions are no longer having one. And it does not forecast upper houses in any jurisdiction.
Where the method rests on an assumption that cannot be tested before the votes are counted, the assumption is named in Assumptions and graded by how well supported it is. Where the current application rests on thin evidence, that is stated prominently on its own page rather than buried.
The model
The model is one probabilistic description of the election, fitted and simulated in three stages, each conditioning on the one before. The first stage estimates each party’s jurisdiction-wide support, day by day, from the polls. The second distributes each party’s jurisdiction-wide support across the individual electorates. The third counts preferences within each electorate to a winner. Uncertainty is carried through by simulation: a single run draws one trajectory of jurisdiction-wide support, one distribution of it across the electorates and one preference count in each, and every published figure is a summary over many thousands of such runs.
Stage one: pooling the polls
The foundation is Simon Jackman’s specification for pooling polls over a campaign (Jackman 2005): a latent daily series that follows a random walk, with each poll observed around that latent state plus a persistent lean for the pollster that published it, which the literature calls a house effect. The extension here is that the state is a vector rather than a single number.
Why a vector, and why log-ratios. Vote shares are compositional: they cannot be negative and they must sum to one. A normal random walk applied directly to shares fails in three ways. It can push a share below zero, it does not conserve the sum, and it treats a one-point move as the same thing whether a party sits at 30 per cent or at 5. The additive log-ratio transform repairs all three (Aitchison 1982). One party is nominated as the reference, and each other party’s coordinate is the logarithm of its share divided by the reference party’s. That produces an unconstrained vector on which Gaussian increments are coherent, and transforming back restores a valid set of shares automatically.
The reference party is a configuration entry, and the choice is a parameterisation rather than a modelling assumption: the reference party is not privileged in the estimates, and all published quantities are back-transformed shares.
The daily increment carries a full covariance matrix rather than one variance per party. This matters for the seat layer: it allows the model to learn that one pair of parties moves together while another pair moves independently, and that structure determines whether errors across electorates compound or cancel.
House effects. Each pollster has a persistent tendency to read parties high or low, which enters additively in log-ratio space. Where a pollster has changed its survey mode over the period, each mode is treated as a separate instrument rather than assuming the house effect carried across the change.
The industry offset. The total error of a final poll has four components: sampling noise, which is symmetric and cured by aggregation; genuine late movement, which is not the pollster’s fault; the house effect, which is that one pollster’s persistent lean; and the industry offset, the amount by which every pollster was wrong together in the same direction. Aggregation cures the first, and estimating each pollster’s lean against the anchoring elections removes the third, so the offset is the component that survives, and it is what was large at the 2019 Australian federal election.
Within one campaign the offset cannot be identified at all. Three pollsters finishing on 31, 32 and 33 for the same party tell you their differences but not where the cluster sits relative to the truth: the common offset and the true level are perfectly confounded, and no volume of polls resolves it. Anchoring the latent series to the known result at past elections breaks that confound. Across several cycles, house effects are separated by variation between pollsters at a given election, and the industry offset by variation across elections. Anchoring a state-space model of Australian polling to known election results is Mansillo and Jackman’s approach in their account of the 2019 failure (Mansillo and Jackman 2020), though they anchor to the two elections bounding one parliamentary term. Extending the anchoring across four cycles in order to identify an industry-wide offset is this site’s own construction, and it is listed as such in what is established and what is ours.
The offset is compositional rather than a single number. A single bias term can represent a shift between two parties and nothing else, which is sufficient only where the contest is genuinely a two-party one. Where three or more parties are in play, an industry-wide error has a direction across the whole composition, and that direction is what determines which candidates reach a final count.
Two disciplines apply to the prior on how much the offset varies between elections. Where the number of anchoring elections is small, the likelihood cannot distinguish an unbiased industry from one whose biases happened to cancel, so the prior has to do that work; left unpenalised, the fit drives the spread to zero and erases the term the model exists to capture. And the prior’s mode must be bounded away from zero, which a half-normal does not achieve, since a half-normal is maximised at the origin and so mildly rewards exactly the collapse being guarded against. A gamma prior with shape above one, whose density is zero at the origin, is used instead, with its rate set by prior-predictive check.
What the anchoring elections identify is the scale of common error, not the direction of the next one. A handful of cycles is too few to establish that the industry leans persistently one way, and a pollster that has missed publicly tends to revise its method, which erodes whatever persistence there was. So the coming election’s offset is modelled as a new draw: centred on no error, with its spread set by how large the common errors have been. A term built this way adds a source of uncertainty without supplying a direction, and adding a source of uncertainty can only widen a coherent range. The consequence is that the offset widens every published range and leaves the central estimates essentially unchanged, and this is checked mechanically within each build by fitting with the term on and off. If the version with the offset were more confident than the version without it, the implementation would be wrong.
Observation model. Poll counts are drawn from a Dirichlet-multinomial distribution centred on the latent shares, with an estimated concentration parameter. That parameter absorbs the overdispersion that a design effect would otherwise be used to assert, so counts enter at nominal sample size and no design effect is applied on top; applying both would double-count the same dispersion. The fitted concentration implies an effective sample size that saturates as nominal size grows, which is the expected behaviour when total survey error exceeds sampling error (Shirani-Mehr et al. 2018).
Campaign variance. A single multiplier scales the daily innovation variance during the formal campaign period, with a prior centred on no change rather than bounded to an increase, so that the data may report a quieter campaign as readily as a noisier one.
Because the random walk is driftless, the central forecast does not trend. Uncertainty grows with the time remaining rather than the central estimate moving towards an expected outcome. A reader looking at a forecast five weeks out is seeing today’s estimated position with the width that five weeks of possible movement implies, not a projection of where opinion is heading.
Stage two: distributing the vote across electorates
Stage one ends with, for each simulation draw, a single jurisdiction-wide share for every party. This stage turns each of those draws into one share per party in every electorate, constrained so that the electorate shares, weighted by votes, average back exactly to the jurisdiction-wide share they came from. Stage one has fixed how much vote each party has; what remains to estimate is where that vote sits.
Parties with local history. Each electorate starts from its first preferences at the previous election. Every party’s electorate shares are then adjusted so that the vote-weighted average across all electorates reproduces that draw’s jurisdiction-wide estimate exactly, for every party. This is iterative proportional fitting, and it is what guarantees that the electorate figures aggregate back to the published jurisdiction-wide figures. A reader can check that with a calculator.
Parties without local history. A party polling far above anything it has previously recorded in the jurisdiction cannot be handled by extrapolating its own local coefficients, because the relationship was never observed at that level of support. The approach is to use the best available adjacent evidence: the same party’s result at a recent election in the same area on comparable geography, shrunk towards what a Census covariate predicts. How far it is shrunk is governed by a reliability parameter estimated in another jurisdiction where the same surge has already been observed and resolved, and transferred with its uncertainty deliberately inflated, because a coefficient from elsewhere is informative about this jurisdiction rather than authoritative over it.
This transfer can fail in three ways. A party moving from a low base to a high one is not the same party expanding within one coalition of voters: the marginal voter at the high level is by definition unlike the loyalist at the low one, so the real profile should be expected to be flatter than the historical gradient, and a proportional-expansion assumption would over-concentrate the vote and overstate seats won. The question of which established party loses those votes is genuinely uncertain and consequential, because it determines the exclusion order and therefore which two candidates reach the final count. And where the surging party campaigns on immigration restriction, its geography is not simply monotone in immigrant population share, since the outer growth corridors of a large city are often simultaneously its most diverse and its fastest-growing parts.
Two variants of the adjustment are run. The primary takes the surging party’s above-average gains disproportionately from one established party, on coefficients measured where the surge has already been observed and weighted by how much vote each party has available to lose. The alternative takes them from every party in proportion to its own vote. Neither is validated, and the difference between them is reported rather than hidden.
Independents. Until nominations close, the independent vote is drawn from the previous election’s field: 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.
Redistributions and new electorates. Starting points are defined on the boundaries the coming election will be fought on. Where those boundaries differ from the previous election’s, each electorate’s starting point is a notional result: the previous election’s votes reassembled onto the new boundaries from polling-place and small-area counts, in the same way federal results are moved onto state boundaries. A newly created electorate has no previous result of its own, so its starting point is entirely notional, and a notional starting point is an estimate rather than an observation. The current application does not exercise this path, because its election is fought on the same boundaries as the previous one.
Stage three: preferences and seats
A full preferential count is run in every electorate, excluding the lowest-polling candidate and transferring their ballots until one candidate holds a majority. Transfer proportions are drawn from priors estimated from past distributions of preferences in the same jurisdiction and from the registered how-to-vote cards.
Every electorate is treated as a general multi-candidate contest, with the classic two-party case emerging as a special case rather than being assumed. That is the point of the architecture: in a chamber where one major party fails to reach the final count across a substantial number of seats, a framework that assumes the two majors contest every seat cannot describe the result. The 2026 Victorian election is a working example of how much turns on this stage. 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, projected 31 Coalition, 30 Labor and 27 One Nation (Raue 2026). Identical first preferences, a difference of 15 seats for one party: the preference model, not the primary vote, decides the chamber.
Where full preferential voting applies, ballots essentially never exhaust and all transferred votes are redistributed among remaining candidates. Jurisdictions where that does not hold are listed in Scope and generality.
What the published ranges include
Carried through to the published plausible ranges: sampling error in each poll; between-pollster differences; error common to the whole polling industry, estimated across the anchoring elections; day-to-day movement in opinion; and uncertainty in the electorate-level distribution, including any cross-jurisdictional transfer it rests on.
Not carried through: the possibility that the relationship between demography and vote is itself changing, which is discussed in Known limitations; and, where the preference layer has not been run, any preference-flow uncertainty.
What the plausible range means
The model is estimated in a Bayesian framework, and the plausible range is a 90 per cent credible interval. That licenses the sentence this site publishes with every figure: the model puts a nine-in-ten chance on the result falling inside the range. A frequentist confidence interval does not license that sentence, although it is routinely read as though it does. A credible interval is a probability statement about the unknown quantity given the data and the priors. A confidence interval is a statement about replicability: how often the procedure that produced it would cover the truth over repeated samples. One speaks about this election; the other speaks about the long-run behaviour of a method.
The width of the range is not itself a consequence of the Bayesian choice. Any method that propagated poll sampling error, house effects, the industry offset, daily movement and the electorate-level transfer jointly would produce a wide range, and most published commentary is narrow because it propagates almost none of them.
Where the Bayesian commitment is load-bearing is narrower and more consequential. The prior on how much the industry offset varies between elections supplies information the likelihood cannot. With a handful of anchoring elections the data cannot distinguish an industry that is unbiased from one whose biases happened to cancel, and no volume of additional polling within a cycle changes that. A prior is the only vehicle for the knowledge that pollsters have been wrong together before. The same holds in the electorate layer, where a hierarchy with few groups at its upper level leaves some variance components weakly identified: a posterior stays honestly wide on a quantity the data do not pin down, where maximum likelihood would sit on a flat ridge or a boundary and report a point estimate anyway. That property is described in Estimation and diagnostics.
One thing the framework does not deliver. Bayesian inference does not make a forecast calibrated. Coherence and calibration are different properties: a posterior can be internally consistent and still assign nine-in-ten ranges that contain the truth far less than nine times in ten. Calibration is an empirical question, settled by scoring forecasts against outcomes out of sample, and what has and has not been tested here is set out in Validation.
Data
What the method requires, in any jurisdiction. The sources used in the current application are listed on the Victorian application page.
Poll series across several cycles. Every published voting-intention poll for the jurisdiction, each carrying its pollster, fieldwork dates, sample size and survey mode. Methodology is always taken from the original publisher rather than from a media summary. Polls are captured with a timestamp and never overwritten, so that a data revision can always be distinguished from a model change.
Several past cycles are needed, not one. The historical cycles are not padding. They are what identifies the industry offset described in the model, and without them that term rests on its prior alone.
Results at each anchoring election. First preferences, two-candidate-preferred counts, distributions of preferences and polling-place results from the electoral commission, for every election the latent series is anchored to.
A recent election on comparable geography. Results from another recent election in the same area, used to set the electorate-level shape. These need moving onto the current electoral boundaries using the commission’s count of votes by Census small area.
Census and boundary files. Census counts at small-area level, and the electoral boundaries the coming election will be fought on. Where a redistribution has occurred, electorate-level comparisons across it are not made without redistributing first.
Structural breaks in the instrument. Any change in what pollsters ask, as distinct from a change in opinion, must be identified and dated. A party added to the prompted list affects every pollster at once, so it is neither drift nor a house effect, and polls before the change cannot inform that party’s series.
Scope and generality
This is a general method rather than a single-state one. Nothing about a jurisdiction is written into the code. Seat counts, party sets, electoral rules, anchoring elections and the mapping of each data source are read from a configuration file, and adding a jurisdiction means adding a configuration rather than altering the model.
That claim is about the architecture, which is the only part of the generality yet built. One jurisdiction has been fitted. The next is New South Wales in 2027, and the working rule is that nothing is generalised further until a second jurisdiction has actually been run.
The method is implemented for full preferential single-member contests, which is the system used in six of the nine Australian lower houses. The remaining three differ in ways that are changes to the method rather than entries in a configuration file.
| Jurisdiction | Lower-house system | Status |
|---|---|---|
| Commonwealth House of Representatives | Full preferential, single member | Supported; not yet applied |
| New South Wales | Optional preferential, single member | Requires exhaustion |
| Victoria | Full preferential, single member | First application |
| Queensland | Full preferential, single member | Supported; not yet applied |
| Western Australia | Full preferential, single member | Supported; not yet applied |
| South Australia | Full preferential, single member | Supported; not yet applied |
| Tasmania | Hare-Clark, five divisions of seven | Requires a quota count |
| Northern Territory | Full preferential, single member | Supported; polling series sparse |
| Australian Capital Territory | Hare-Clark, five electorates of five | Requires a quota count |
Electoral systems are as at the date of this version. Upper houses are out of scope in every jurisdiction; a separate model for upper houses is being investigated, and nothing on this site rests on it.
Exhaustion. The preference layer assumes every ballot is numbered through to a final pair. That holds where full numbering is compulsory. It does not hold under the optional preferential voting used in New South Wales, nor under Hare-Clark in Tasmania and the Australian Capital Territory: a Tasmanian ballot must number at least as many squares as there are seats and may stop there, and an ACT ballot is formal from a single first preference, so preferences exhaust in both. Exhaustion alters the running total against which exclusion and election thresholds are computed, so it is a change to the count itself rather than a setting applied to it.
Multi-member counting. Tasmania and the Australian Capital Territory elect several members in each division by quota, transferring surpluses at a fraction of a vote. The single-member architecture here does not contain that as a special case, in the way an ordinary two-candidate contest falls out of the general preferential count. It needs a separate count engine, and that engine does not yet exist.
Robson Rotation. Tasmania and the Australian Capital Territory rotate candidate order across ballot papers so that no candidate holds the favoured positions, and a party’s capacity to direct preferences is correspondingly weak. The how-to-vote-card component of the seat layer therefore has no analogue in those jurisdictions rather than a different parameterisation.
There is a fourth limit, and it is statistical rather than structural. The industry offset described in the model is identified only by anchoring the latent series to known results across several past elections in the same jurisdiction. Where a jurisdiction has few anchors the prior is already carrying work the likelihood cannot, and Tasmania, the Northern Territory and the Australian Capital Territory have little or no published voting-intention series at all. There the offset would rest on its prior alone, which is precisely the failure that prior exists to prevent. The method’s reach is bounded by where polling exists, not by the code.
Assumptions
Each assumption is graded by how well supported it is. Validated means tested out of sample. Supported means consistent with evidence but not directly tested. Reasoned means argued from principle. Assumed means taken as given.
These are the assumptions the method makes wherever it is applied. The further assumptions made by the current application are graded on the Victorian application page.
| # | Assumption | Grade |
|---|---|---|
| M1 | Vote shares are compositional and best modelled in log-ratio space | Established in the literature |
| M2 | Latent opinion follows a driftless random walk, so the central forecast is flat and uncertainty grows with time remaining | Supported |
| M3 | Each pollster has a persistent lean, separable from the industry-wide offset given multiple anchoring elections | Established in the literature |
| M4 | The scale of the industry offset is exchangeable across cycles: the coming election’s offset is a new draw, centred on no persistent direction, with spread estimated from past cycles | Reasoned |
| M5 | The relationship between demography and vote is stable over the estimation period | Assumed; see Known limitations |
| M6 | Where a party has no local history, its geography can be built from the same area’s most recent comparable election, with the shrinkage calibrated in another jurisdiction where the surge has resolved, given inflated uncertainty | Reasoned |
Estimation and diagnostics
The model is built in the R statistical environment, with the probabilistic core written in Stan. The aggregator is fitted with Hamiltonian Monte Carlo via cmdstanr, in four chains, and the electorate-level model is a separate Stan program fitted the same way. An independent Python implementation is planned, to validate the R build by reproducing its results from the same data; it does not yet exist. Convergence is checked on R-hat, bulk and tail effective sample size, divergent transitions and energy diagnostics, and a fit is not used for publication unless the quantities that enter the published figures are well mixed.
An identification result worth stating plainly. In the electorate-level model, the total variance of the electorate effects is sharply estimated. How that total divides between two components, and the reliability weight governing how far each electorate is pulled towards its Census expectation, are not: the data constrain the total and say little about the split. This is a real property of a hierarchy whose upper level has only two groups, not a defect of the software, and no amount of sampler tuning changes it. It does not affect the published estimates, which depend on the total rather than on the split, and the electorate effects themselves mix well. It is recorded here because a reader checking the diagnostics would otherwise find it unannounced.
Each published run is identified by a tag recorded with every figure, and each run is logged with its model version, configuration, date and the reason it was made.
What is published
The model publishes three distinct results, and they answer three different questions. Reading one result as the answer to another question is the most common way to misread a forecast.
Jurisdiction-wide first-preference share. For each party, its projected share of the formal first-preference vote across the whole jurisdiction. Reported as a median and a plausible range.
Electorate first-preference share. The same quantity within each electorate. Reported as a median and a plausible range. These are estimates of the vote, not of the winner.
Probability of the highest first-preference vote. For each electorate, the proportion of the model’s estimates in which a given party has more first preferences than any other. This is not a probability of winning the seat. Australian lower-house elections are decided after preferences are distributed, and a candidate can lead the first count and lose the seat.
Win probabilities and seat totals are published only once the candidate field and the parties’ how-to-vote cards are known. In the current application they are not yet published; see Known limitations for why not before.
Validation
A model earns trust through tests, and different tests earn different amounts of it. This section lists every test the model has faced and what each can establish, and then states which parts have faced no test at all. The weight a reader should put on each published figure follows from where the part of the model that produced it sits on this list. The tests described are those of the current application.
In-sample checks, passed. The fitted model recovers the known results at each anchoring election to within a fraction of a point. Calibration widens rather than narrows the ranges, which is the direction the method requires.
Out-of-sample, one test. An earlier and simpler version of the electorate-level method was tested against the 2026 South Australian result; the test and its figures are reported on the Victorian application page. That is one election. The version in current use builds on the one tested, and cannot itself be tested on South Australia, because South Australia now informs its priors. Its first genuine out-of-sample test will be the Victorian result.
Pre-registered, not yet run. A leave-one-election-out validation protocol was written and locked, with a recorded hash, before the data that will test it were examined. It holds out each past election in turn, refits, and scores the forecast against the held-out result. That confirmatory run has not been made. Until it has, a model that samples cleanly and recovers its anchors is a step rather than a validated result, and nothing on this site should be read as a validated forecast.
Not validated at all. The preference and seat layer, which has not been run end to end.
Results, once there are any, appear on the Track Record page.
Known limitations
This section records what the model cannot do and where it is most exposed, even when every diagnostic passes. Each entry names the limitation, the reason it exists, and what it means for a reader of the published figures. These are the things the model’s own ranges do not fully capture, which is why they are written down rather than left to the arithmetic.
A party without local history is always the largest exposure. Where the model carries a party polling far above anything it has recorded locally, that party’s figure has the widest range and is the most exposed to error. Every electorate estimate inherits the jurisdiction-wide level, so if that level is several points out, every electorate moves with it. In the current application this is One Nation; three of the four assumptions behind it rest on one South Australian election, and the fourth on a candidate list that does not yet exist.
Polled support is not the vote. The aggregator estimates the support pollsters measure. A party’s actual vote also depends on how many electorates it contests. Treating polled support as the vote assumes a full slate, and where a party has not confirmed one this overstates it to a degree that cannot be quantified in advance.
No seat forecasts before the how-to-vote cards. Preferences decide these elections; preference flows depend on who is standing and what the parties recommend; neither is known until nominations close and the cards are registered. Any seat forecast published before that point projects a candidate field that does not exist.
Realignment is the standing vulnerability. The model assumes the relationship between demography and vote is stable. Gethin, Martínez-Toledano and Piketty document a reversal of the education cleavage across 21 Western democracies: voters with higher education moved from being less likely to support left parties in the 1950s and 1960s to about ten percentage points more likely by 2016 to 2020 (Gethin et al. 2022). A minor party reaching a quarter of the vote is on its face a realignment event rather than movement along a stable gradient, and a model fitted on earlier cycles may encode a relationship the coming election is breaking. This is a reason to widen the uncertainty on the electorate layer considerably rather than to abandon it, and it is stated here because a reviewer would otherwise identify it first.
Threshold counts are fragile. A count of how many electorates a party leads is a threshold applied to a continuous quantity. Electorates sitting near an even chance can change sides between two runs of the same model on nothing but simulation noise, and they do. Where such a count is published, the electorates on the threshold are named.
The electorate layer is narrower than originally specified. The design calls for a booth-level model on several covariates; what exists is an electorate-level model on one covariate, for the party without local history. The narrower version supports statements about electorates rather than about people.
What is established, what is standard, and what is ours
Established in the literature. The latent-state formulation with house effects (Jackman 2005). Log-ratio transforms for compositional data (Aitchison 1982). Multilevel regression with poststratification (Park et al. 2004) and its dynamic form (Linzer 2013). The finding that total survey error substantially exceeds sampling error (Shirani-Mehr et al. 2018). Anchoring a state-space model of Australian polling to known election results, and the diagnosis of herding among Australian pollsters (Mansillo and Jackman 2020). The reversal of the education cleavage across Western democracies (Gethin et al. 2022).
Standard practice among Australian practitioners. Poll aggregation with house-effect correction. Post-election scoring of pollsters. The Mackerras pendulum as a seat-projection baseline.
Ours, and untested. The particular assembly of the pieces above is this site’s own construction, and four choices define it. Primary votes are treated compositionally from the polls through to the seat count, so every figure respects the constraint that shares are non-negative and sum to one. The industry offset is estimated by anchoring to four past elections rather than the two bounding a single term, with its between-election variance deliberately prevented from collapsing to zero, so the possibility of a common polling error survives into every published range. The published Australian treatments anchor within one term and estimate pollster-specific house effects; extending the anchoring across cycles to identify an offset common to the whole industry is ours. A party with no local history is calibrated from an adjacent jurisdiction where the same surge has already been observed and resolved, with the transfer’s uncertainty inflated. And the seat architecture is crossbench-first: every electorate is a general multi-candidate contest, with the familiar two-party contest recovered as a special case rather than assumed. Each choice has precedent of its own. The assembly of the four does not. This combination is a proposal, not an established method. It has not yet been validated against an election result.
The generalisation is itself untested. The architecture is jurisdiction-neutral by construction, but a configuration-driven design is only demonstrated to be general once a second jurisdiction has been fitted, and none has. Until then the claim in Scope and generality is a claim about how the code is written rather than a result.
Availability
Published with each release: the full first-preference estimates for every party in every electorate, with ranges, as a CSV file, free to use with attribution. Every figure names the model run that produced it.
Not currently published: the model source code and the poll database. The data the method requires are described in Data, the sources used now are named on the Victorian application page, and all are publicly available from the electoral commissions, the Australian Bureau of Statistics and the pollsters.
Change log
Version 1.0, 27 October 2026. First publication, covering the aggregator, the electorate-level projection and the data behind them. The preference and seat layer is described as designed but not yet implemented. No previously published figures are affected, as this is the first release.
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 = {The 98th {Percentile} Election Forecasting Model},
date = {2026-10-27},
url = {https://the98thpercentile.pages.dev/methodology.html},
langid = {en-AU}
}