How the model works
The simulator uses a hierarchy of nine rules to estimate preference flows for each elimination round. Rules are tried in order from highest to lowest confidence; the first one that produces a valid result is used. Each round in the output shows which rule fired and why.
The rule hierarchy
The key insight behind the model is that preference flows depend on who is eliminated and who remains — not just the eliminated candidate's classification in isolation. A Greens voter behaves differently when choosing between Labor and One Nation than when choosing between Labor and a teal independent. All nine rules account for the full composition of the remaining field.
An exact match exists in the 2025 AEC distribution of preferences data for this specific combination of eliminated classification and remaining classifications, with five or more observed instances. The model averages the actual preference flows across all matching observations.
133 observed combinations underpin this rule. The most common examples: Greens eliminated with [Coalition, Labor] remaining (n=72, 81.5% to Labor); One Nation eliminated with [Coalition, Greens, Labor] remaining (n=69, 61.9% to Coalition).
The threshold is n≥5, not n≥1, because single observations can reflect idiosyncratic seat dynamics rather than general patterns. Confidence: 78–95% depending on sample size and variance.
When the 2025 federal data has no high-confidence match, high-quality data from another election provides a better estimate. The current calibrations use the SA 2026 state election preference flows (Bonham, April 2026).
Three scenarios are calibrated: Coalition eliminated with [Labor, One Nation] remaining (33.9% to Labor — the single 2025 federal observation of 80% was treated as unreliable); Greens eliminated with [Labor, One Nation] remaining (84.8% to Labor); and Other Unaligned eliminated with [Labor, One Nation] remaining (55.9% to Labor).
Confidence: 70%
No high-confidence exact match exists, but a closely similar observed scenario does. Similarity is calculated as the Jaccard index on the multisets of remaining classifications — intersection divided by union. A 4-candidate remaining field that differs from an observed scenario by one candidate scores 4/5 = 80%.
A coverage check is applied: if the nearest match's flows don't include all remaining classifications, the match is skipped to prevent silently assigning 0% to a candidate.
Confidence: 52–75% depending on similarity score and sample size.
No direct or near match is available. The model tries substituting ideologically equivalent classifications: Labor ↔ Greens (left-bloc swap), Coalition ↔ One Nation (right-bloc swap), or both simultaneously. Flows from the swapped scenario are found and the labels are translated back.
Example: Coalition eliminated with [Greens, Labor, One Nation] remaining has no direct data. A right-bloc swap looks up One Nation eliminated with [Coalition, Greens, Labor] remaining (n=69) and re-labels Coalition→One Nation. Result: One Nation ~62%, Greens ~20%, Labor ~18%.
Confidence: 45%. The left-bloc assumption (Labor ≈ Greens voters) is stronger than the right-bloc (Coalition ≈ One Nation voters).
All swap attempts failed. For Other Left-leaning candidates, the model uses Labor's observed flows as a proxy (falling back to Greens). For Other Right-leaning, it uses Coalition's flows (falling back to One Nation).
Confidence: 30–35%. Other Left-leaning covers a very wide range — Climate 200 teals and pro-Palestinian community candidates behave quite differently. Flag results under this rule for manual review.
Only for Other Unaligned eliminated candidates. The model runs Rules 1–5 independently for Other Left-leaning and Other Right-leaning as proxies, then averages the two resulting distributions.
Confidence: 30%. The averaging can overstate One Nation's share in One Nation-heavy scenarios; treat with caution.
All data-based rules have failed, but at least one remaining candidate shares the eliminated candidate's classification. 80% of votes go to same-classification candidates collectively; 20% is split equally across other unique remaining classifications. The 80% is calibrated against observed Coalition-to-Coalition flows.
Confidence: 25%.
All rules above have failed with no same-class candidates remaining. Votes are distributed by ideological bloc — Left (Labor, Greens, Other Left-leaning), Right (Coalition, One Nation, Other Right-leaning), and Unaligned — with calibrated weights favouring ideological proximity.
Confidence: 20%. This rule never fired in the 150-seat 2025 backtest. It is a structural safety net for exotic scenarios.
Absolute last resort. No match of any kind available. Votes are distributed equally across all unique remaining classifications, then proportionally within each class by vote totals.
Confidence: 10%. This rule never fired in the 150-seat 2025 backtest. Any output under this rule should be manually reviewed and overridden.
Reliability rating
Each simulation produces a confidence rating based on three factors: the quality of data used across all elimination rounds, how close the critical eliminations were, and how decisive the final margin is. These are combined into a single score; the thresholds are calibrated against the 2025 backtest results.
A vote-weighted average of rule quality scores across all elimination rounds. Rounds with more votes have more influence. An elimination round using Rule 1 with n≥30 observations contributes at 95% confidence; one using Rule 5 (proxy) contributes at 35%.
The penultimate elimination — the one that determines which two candidates make the final — is the critical round. If the margin between 2nd and 3rd place at that point is very small (under 1.5% of total votes), the predicted final pair is highly uncertain and the score is penalised accordingly. Earlier tight eliminations are also flagged as they can cause cascading errors.
A predicted result of 50.4% vs 49.6% is directionally unreliable even if the data quality is high. A small predicted margin reduces the rating.
Data sources
The primary data source is the AEC's HouseDopByDivisionDownload-31496.csv file from the 2025 Australian federal election. This contains the full distribution of preferences for every House of Representatives seat, round by round. 133 distinct observed preference flow combinations were extracted from this data and underpin Rule 1.
Three specific scenarios lack reliable 2025 federal data and are instead estimated using preference flow data from the 2026 South Australian state election, published by Kevin Bonham in April 2026. The SA data is preferred for these scenarios because the single available 2025 federal observation was judged unreliable (n=1 in an unusual seat context).
Candidate classifications for the 2025 backtest were assigned manually using Climate 200 endorsement lists, AEC TCP flow data, candidate backgrounds, and the Bonham psephology blog classification framework. Classifications for new candidates must be assigned by the user based on the same criteria.
Known limitations
These are the most significant sources of error, based on analysis of the 2025 backtest deviations. All large deviations in the backtest occurred in seats that match one or more of these scenarios.
The biggest source of error. "Other, Left-leaning" covers both pro-Palestinian community candidates (whose voters give ~84% to Labor in a Labor vs Coalition final) and Climate 200 teals (whose voters give ~50% to Labor). The model uses the average (~67%) and cannot distinguish between them without additional input.
In established teal-incumbent seats (Mayo, Mackellar, Wentworth), Coalition voters send 80–85% to the teal independent in a [Labor, Other Left-leaning] final — significantly higher than the model's average of ~67%. The model systematically underestimates the teal candidate's share in these seats.
In rural regional seats (Maranoa, Cowper), minor right-wing candidates send 55–60% to One Nation, versus the model's average of ~37–43%. The model underestimates One Nation's share in these contexts.
The model was developed using the 2025 data and then tested on the same data. This overstates real-world performance. Out-of-sample performance (future elections) will differ, particularly if the political landscape changes significantly — new parties emerge, voter behaviour shifts, or the candidate field has configurations not well-represented in 2025.
The simulator models what happens to votes after the count begins. It does not predict what the primary votes will be. Prediction of primary vote shares is a separate problem — and the uncertainty in that prediction is not captured in the reliability rating.