Before any preference can be simulated, any swing can be modelled, or any demographic profile can be built, every candidate in the field needs to be assigned a “classification”. This is the base design decision underpinning The Swing Is On.
These groupings are more than labels – they reflect something real and measurable: candidates within the same group send their voters’ preferences in roughly the same direction. And it sets us up for the usability of the model in future elections with a different set of candidates.
That distinction matters because, under Australia’s compulsory preferential voting system, what happens to a candidate’s votes when they’re eliminated shapes the final result just as much as their primary vote does.
Why seven classifications?
The Swing Is On has seven candidate classifications. The first four of these are:
- Labor – the Australian Labor Party
- Coalition – the Liberal Party, National Party, Liberal National Party in Queensland, and Country Liberal Party in the Northern Territory. While four technically separate entities, these parties have had a long-standing (though more recently on-and-off-again) relationship at the federal level, and don’t often contest seats against each other (only 7 out of 150 in the 2025 federal election). Their voters’ preferences generally flow similarly.
- Greens – the Australian Greens
- One Nation – Pauline Hanson’s One Nation
Those four are straightforward: party membership determines the classification.
The remaining three cover everyone else – minor parties and independents alike:
- Other, Left-leaning includes independents and candidates belonging to minor parties with more left-leaning values (on social issues, climate change, immigration), those backed by Climate 200 and/or Muslim Votes Matter, and – in uncertain cases – candidates where 65% or more of their two-candidate-preferred preferences flowed to a left-leaning candidate over a right-leaning candidate. In 2025, this group most notably included the “teal” independents, the Animal Justice Party, and the Legalise Cannabis Party.
- Other, Right-leaning conversely includes independents and candidates belonging to minor parties with more right-leaning values, and – in uncertain cases – candidates where 65% or more of their two-candidate-preferred preferences flowed to a right-leaning candidate over a left-leaning candidate. In 2025, this group most notably included Trumpet of Patriots, Family First, and the Libertarian Party.
- Other, Unaligned covers candidates with no consistent ideological positioning – those whose preference flows showed no clear directional pattern and for whom limited publicly available information existed on their political beliefs.
That decision to split “Other” into three buckets is reasonably novel as far as I’m aware – but one I feel is key in a landscape where there is now a historically large crossbench in the House of Representatives, and a combined major party (Labor and Coalition) vote falling under 50%.
I feel this approach strikes the right balance between not over-complicating the model with dozens of classifications for a series of micro-parties (many with little political longevity), and avoiding treating this entire cohort as an unimportant monolith – which is far from reality.
However, there are limitations to this approach that are worth bearing in mind. In particular, it creates a certain degree of design bias – though every meaningful political model requires some difficult choices – and it also creates some unlikely friends.
The trade-off in the “Other” categories
Every classification involves a trade-off between precision and tractability. The “Other” categories are where that trade-off is most visible.
A Climate 200-backed teal independent in a wealthy (traditionally Coalition-held) seat, a Socialist Alliance candidate running in a younger inner-metro seat, and a community-based rural independent in regional Queensland might all get classified as Other, Left-leaning – but their voters send preferences quite differently.
In 2025, the model slightly over-estimates how strongly progressive preferences flowed to rural-seat independents compared to their urban counterparts, which contributed to the only two incorrect winner predictions for 2025 in the model: Cowper and Forrest.
I’ve published a separate article exploring The Swing Is On’s preference distribution model, including these marginal errors, in greater detail.
The groupings exist because the aggregate difference in preference behaviour between left-leaning, right-leaning, and unaligned “Others” is real and consistent enough to be worth capturing. Within each group, individual candidates will vary. The model uses group averages, which are the best available estimates – but they will not perfectly represent any single candidate.
This is why the preference simulator shows a reliability rating alongside every result, and why that rating is lower in seats where unusual “Other” candidates dominate the field.
As I’d recommend with any seat prediction model, take any non-traditional seat (particularly one with a small existing dataset from similar seats) with a grain of salt. Psephology isn’t magic after all.
