In most countries, the candidate with the most votes on election night wins. Australia does it a little differently.
Under compulsory preferential voting, every valid ballot paper eventually ends up counted for one of the final two candidates standing – which means every elimination round matters, and where the preferences go at each step can determine the result.
The Swing Is On’s preference simulator models that process for any candidate field you give it. You enter the candidates, their classifications, and their primary vote counts, and the tool works through each elimination in order, distributing votes and tracking who is left. The output is a predicted two-candidate-preferred (2CP) result, along with a reliability rating that reflects how confident the model is in that prediction.
The key insight
Where a candidate’s votes go when they’re eliminated depends on who is still in the race. A Greens voter who has to choose between Labor and Coalition will behave differently from a Greens voter choosing between Labor, Coalition, and One Nation. Both involve a Greens elimination, but the remaining field changes the available options.
The simulator accounts for this. Before distributing any round’s votes, it identifies the specific combination of: who was just eliminated, and who is still standing. That combination is the key to finding the right data.
A hierarchy of twelve rules
The model works through a ranked list of twelve rules, stopping at the first one that produces a valid result. Rules at the top are data-driven and specific. Rules toward the bottom are progressively more approximate, used only when more precise options are exhausted.
Rule 1 – direct observation: The model searches 133 combinations from the 2025 AEC distribution of preferences data for an exact match with five or more real examples. If found, the observed flows from those actual elections are applied directly. The Greens→[Labor, Coalition] scenario has 72 observed instances from 2025, giving a flow of 81.5% to Labor. One Nation→[Coalition, Greens, Labor] has 69 observations, with 61.9% to Coalition.
The five-observation threshold exists because fewer instances can reflect unusual seat dynamics rather than general patterns. A Greens elimination in a seat with a particular local personality contest might produce an atypical result. Five instances, averaged together, smooth out that noise.
Rules 2 to 6 – closest available data: If no exact match exists at five or more observations, the model searches for the closest structural equivalent – first among higher-confidence matches, then progressively through smaller sample sizes. The similarity measure used (the Jaccard index) compares the remaining fields of two scenarios to find the best structural match. A scenario with four of five matching candidates might produce a reliable enough estimate; a scenario with only two of five might produce a less reliable one. A coverage threshold also applies: the matched scenario must direct at least 50% of its flow to classifications present in the current field.
Rule 7 – bloc-equivalent swap: If no direct data exists at any sample size, the model tries an informed substitution. Labor and Greens are treated as interchangeable for lookup purposes, as are Coalition and One Nation. A scenario with no direct data often has a close equivalent where the bloc partners are swapped. Coalition eliminated in a [Greens, Labor, One Nation] field has no direct data, but One Nation eliminated in a [Coalition, Greens, Labor] field has 69 observations. This swap yields approximately 62% to One Nation, 20% to Greens, 18% to Labor.
Rules 8 and 9 – ideological proxies: For candidates in the “Other” categories, the model substitutes flows from their closest ideological equivalent. Other, Left-leaning uses Labor’s observed flows as a proxy, with a fallback to Greens. Other, Right-leaning uses Coalition’s flows, with a fallback to One Nation. For Other, Unaligned candidates, the model runs the proxy logic for both left-leaning and right-leaning equivalents and averages the two distributions.
Rules 10, 11, and 12 – structural fallbacks: When all data-based approaches fail, the model falls back on general principles: same-party loyalty (80% of votes to same-classification candidates), ideological bloc alignment, and as a last resort, equal distribution across the remaining field. These rules are safety nets. In the 150-seat backtest of the 2025 federal election, the bloc alignment and equal split rules never fired.
A note on SA 2026 calibration
Six specific scenarios lack sufficient 2025 federal data but have reliable preference flow data from the 2026 South Australian state election (computed from official SAEC distribution of preferences data). These are applied as hard overrides rather than passing through the normal hierarchy:
- Coalition eliminated with [Labor, One Nation] remaining: 33.4% to Labor, 66.6% to One Nation
- Greens eliminated with [Labor, One Nation] remaining: 80.4% to Labor, 19.6% to One Nation
- Labor eliminated with [Coalition, One Nation] remaining: 71.5% to Coalition, 28.5% to One Nation
- Coalition eliminated with [Greens, Labor, One Nation] remaining: 18.9% to Greens, 28.2% to Labor, 52.9% to One Nation
- Labor eliminated with [Coalition, One Nation, Other Left-leaning] remaining: 19.1% to Coalition, 13.2% to One Nation, 67.7% to Other, Left-leaning
- Labor eliminated with [Coalition, One Nation, Other Right-leaning] remaining: 32.5% to Coalition, 21.5% to One Nation, 46.0% to Other, Right-leaning
One Nation’s national primary vote share has grown substantially since 2025, making these scenarios meaningfully more common in current polling environments. The SA data provides a better basis for modelling them than the thin 2025 federal record. The only other significant comparable election with compulsory preferential voting before the 2027–28 federal election is the Victorian state election this November – when the preference flow data is finalised by the VEC, I’ll consider it for integration.
Consistency with the 2025 data
Running the model back over all 150 House of Representatives seats from the 2025 federal election – using the actual primary votes as inputs – shows how consistently it produces results that match the preference flows it was trained on. Because the model is calibrated on the same data, this is an internal consistency check rather than an independent accuracy test.
| Metric | Result |
|---|---|
| Correct winner called | 148 / 150 seats (98.7%) |
| Correct final two candidates identified | 144 / 150 seats (96.0%) |
| Correct winner within correctly-paired seats | 143 / 144 seats (99.3%) |
| 2CP within ±1pp of actual result | 79 / 144 correctly-paired seats (55%) |
| 2CP within ±2pp of actual result | 114 / 144 correctly-paired seats (79%) |
| 2CP within ±5pp of actual result | 144 / 144 correctly-paired seats (100%) |
| Average 2CP absolute error (correct pairs) | 1.19pp |
The two incorrect winner predictions – Cowper and Forrest – were both extremely tight contests rated Low confidence, and both involved Other, Left-leaning independents in rural or regional seats:
- Cowper – the model correctly calls the final two as the Coalition’s Pat Conaghan and Climate 200-backed “teal” independent Caz Heise. However, Labor’s vote at the last elimination count is modelled to have a 75.8% split to Heise (based on 12 similar instances), but in reality only split 64.6% to Heise – overcompensating her by 2,000+ votes and incorrectly calling her (not Conaghan) as the winner.
- Forrest – this was a very tight three-cornered contest in reality, and Climate 200-backed “teal” independent Georgia Beardman edges out Labor’s Tabitha Dowding by just 144 votes at the final elimination in the model (versus a Labor lead of 758 at this same stage in the real count). In the two preceding eliminations (Greens, then One Nation), Dowding overperformed and Beardman underperformed, leading to a second-and-third place switch. With the incorrect final pair, the model then expects Beardman to do better off Labor preferences (75.8% vs Coalition) than if the situation were reversed (67.0% to Labor vs Coalition). So we end up getting the actual third-placed candidate (Beardman) modelled to defeat the actual winner (the Coalition’s Ben Small).
There is a minor systematic tendency to underestimate the winner’s margin by about half a percentage point on average. The observed preference flows are cross-seat averages: safe seats tend to produce stronger-than-average flows to the winner, while marginal seats produce weaker ones. When the model applies the average flow uniformly, it systematically underestimates comfortable winners. Because Australian federal seats skew towards comfortable wins rather than marginal contests, this produces a small but consistent negative bias – a form of regression to the mean built into the preference flow averaging process.
What the figures confirm is that the model is internally consistent with the data it was built from – it doesn’t contradict itself or produce results that systematically diverge from the observed preference flows. Whether that translates to the same level of accuracy on a future election is a different question, and one only time and new data can answer.
The reliability rating
Every simulation produces a High, Medium, or Low confidence rating. This combines three things: the quality of the rules that fired in each round (weighted by the size of each elimination), how tight the margins were at key decision points, and how comfortable the predicted final margin is.
In the 2025 backtest, High-confidence simulations got the correct pair and winner in all 98 cases. Medium-confidence simulations got the pair right in 37 of 40 and the winner right in all 40. Low-confidence simulations, where the model flagged real uncertainty, got the pair right in 9 of 12 and the winner in 10 of 12 – those two being Cowper and Forrest.
The rating is displayed alongside every round’s output in the simulator. The rule badge shown in each round tells you which rule fired and gives a brief note on why. Checking these is worthwhile in any scenario where the reliability rating is Medium or Low.
What to watch for
The twelve-rule hierarchy covers the breadth of scenarios that actually played out in the 2025 Australian federal election. Where results warrant extra scrutiny is in seats with unusual independent or micro-party candidates – particularly new candidacies with no established preference pattern.
The “Other” candidate classifications are where the greatest errors are likely to be seen. The model uses the group average, which is the best available estimate from the data – but treating the result from a Low-confidence simulation involving an unusual independent as a precise prediction would be a mistake. The reliability rating and rule badges exist precisely to flag these situations.
