A polling figure on its own doesn’t tell you much about seats.
Knowing that the Coalition’s primary vote has fallen five percentage points nationally is interesting. Knowing which of the 150 electorates are affected, and by how much, and what that means for which candidates survive through to the final two – that’s where an election is actually decided.
The swing model on The Swing Is On translates primary vote changes into seat-level outcomes. You specify a set of swing rules describing how you think votes have shifted (or based on actual polling data), and the model applies those rules across the 2025 primary vote in all 150 House of Representatives seats – then runs the full preference simulator on the result to produce a modelled national seat count.
Starting from 2025
The baseline for every swing is the actual 2025 primary vote result in each electorate. Swing rules describe movements on top of that baseline. The model answers a specific question: if 2025 voters had shifted in the way you’ve described, what would the seat count look like?
This framing matters. The model does not forecast what primary votes will be at a future election – that requires judgement about campaign dynamics, leader effects, economic conditions, and other factors the model cannot observe. The swing model’s job is to translate a set of assumed vote shifts into a seat count, efficiently and consistently, across all 150 electorates at once.
What a swing rule contains
Each rule specifies five things: the percentage of voters to shift, which demographic group to target (or all voters), which seats to apply the rule to, which party’s votes to take from, and which party to give them to.
When “all voters” is selected, the arithmetic is straightforward. A rule saying “5% of the Coalition’s 2025 voters in Victoria switch to Labor” takes 5% of Coalition’s vote in each Victorian seat and moves it across.
The geographic scope can be narrowed to a specific state or territory, or to a geography type – inner metropolitan, outer metropolitan, provincial, or rural. This lets you model patterns that are regionally concentrated, rather than applying a uniform national swing that may obscure where the movement is actually occurring.
Demographic targeting
The more distinctive feature of the swing model is the ability to target a specific demographic group. A rule might say “15% of 18–34 year old Coalition voters switch to the Greens” or “8% of renters who voted Labor switch to One Nation.” Applying that rule requires knowing how many of each party’s actual voters, in each seat, belong to that demographic.
For each electorate, the model knows what share of enrolled voters fit the selected demographic, via 2021 Census data using up-to-date Commonwealth Electoral Division boundaries. It also has an estimated probability that a voter in that demographic, in that location, voted for the source party based on booth-level data (more on this in another article).
Multiplying those together gives an estimate of how many votes in that seat are both in the demographic and voted for the source party. The swing rule then shifts the specified percentage of those votes.
This matters because demographic groups are not evenly spread across electorates. Greens voters skew substantially younger than the electorate as a whole, and that skew varies by seat. Applying a swing among young Greens voters needs to reflect that concentration.
Note that this definition of “swing” is a little different to how swings have traditionally been considered in Australian elections – and intentionally so. For all of the models in The Swing Is On, a swing is specifically on primary votes and as a percentage of the votes received for the source party (not the total population). If Labor recorded 30% at the last election, a 10% swing to the Greens transfers 3% of the national vote value (10% of the voters who backed Labor last time), not a full one-third of their voters.
For the first time in a century in Australia, we appear to be on the cusp of an electoral map no longer primarily defined by Labor and the Coalition. Evolving our electoral language is an important part of keeping up with that change.
Multiple overlapping rules
When several swing rules target the same source party in the same seat, the model ensures they don’t collectively shift more votes than the party actually has. Rather than simply adding up each rule’s share and hoping it doesn’t exceed 100%, the model uses a compound formula: each rule operates on what remains after the previous rules, not on the original total.
The total fraction shifted is calculated as one minus the product of what each rule leaves behind. So two rules that each shift 50% of a party’s votes don’t shift 100% combined – they shift 75%, because the second rule takes half of the remaining half. The shifted votes are then distributed across destination parties in proportion to each rule’s individual share.
Working backwards from poll figures
Beyond building swing rules manually, the model includes a second mode: you enter a target primary vote distribution and the model works out which combination of swings most plausibly produced it.
The process starts with a baseline: the estimated primary vote for your chosen scope under 2025 conditions. For “all voters” nationally, the baseline is the population-weighted average of actual 2025 primary vote shares. When a demographic group is selected, the baseline draws on location-calibrated demographic rates for that group.
From there, each classification is identified as a gainer or a loser relative to the baseline. The model then needs to decide how to distribute each loser’s decline across multiple gainers – and it uses preference flow data as a guide. For each loser-gainer pair, it calculates how often, in observed 2025 AEC data, voters who had previously backed the loser ended up with the gainer when the loser was eliminated. This gives a sense of the natural gravitational pull between more aligned ideological blocs (like Labor to Greens voters, and Coalition to One Nation voters).
These propensities seed a transfer matrix that must satisfy two constraints simultaneously: each loser’s row must sum to its total decline, and each gainer’s column must sum to its total gain. Starting from an initial estimate based on the propensities, the algorithm alternately rescales rows and columns until both constraints are met. This technique, known as the Sinkhorn algorithm, produces the transfer matrix most consistent with the preference-flow prior that also exactly accounts for the observed gains and losses.
The result is converted into a set of swing cards – one per loser-gainer pair – in the same format as any manually specified rule. These can then be adjusted, extended, or overridden.
This swing estimates feature can also be used directly within the Track a Poll model, using current (or past) polling averages to simulate how an election would have played out with the same candidates as the 2025 federal election.
What the model assumes and where it may be wrong
The model holds preference flows constant even when the composition of a party’s primary vote changes. A swing that shifts moderate Liberal voters to Labor will produce different downstream preference patterns than the 2025 Coalition average – but the model applies the 2025 flows regardless. If you have reason to think preferences will behave differently under your assumed scenario, the preference simulator supports manual overrides.
The model also assumes that a swing rule specified as “5% of a party’s voters” refers to 5% of their 2025 primary votes in each seat. In a seat where the source party started with very few votes, the absolute number of votes shifted will be small regardless of the percentage. The geographic scope filter helps here – limiting a swing to seats where the source party has meaningful vote share – but the user’s judgement about which rules make sense for which contests remains the crucial input.
The swing model is a tool for structured scenario analysis, not a forecast. Its value is in applying a consistent set of assumed movements across all 150 seats simultaneously, and in translating those movements through the full preference distribution to a seat count. What those movements should be is the question it cannot answer.
