Forecast

Will this government be re-elected?

A statistical model trained on the polling record of every completed federal term since 1974 reads the signature of Albanese’s second term, about 17 months in — as at 16 October 2026.

What this number is

The number above is not a poll of voters and it is not a betting price. It comes from a model that was shown the polling history of every completed federal parliamentary term since 1974 — nineteen terms, thirteen of which ended in the government’s re-election and six in its defeat — and taught to recognise the signature a term shows part-way through: how far the vote has faded from the last election result, and how old the government is. Read against that record, the signature of Albanese’s second term is the signature of a government that goes on to survive.

The honest part of the answer is the width of the band. Nineteen terms and six defeats is still a thin history, so the genuine range runs from an uncomfortable 47 to a comfortable 88 per cent. What the record can say clearly is which company this term keeps: at the same age this term reads like the terms that survived rather than the six that fell.

How the model reads this term

Three readings, all knowable from published polling, drive the call — extended by two the model derives itself: the fraction of the term elapsed, and the interaction of the primary fade with that elapsed fraction.

• Primary vote, −6.7 on 2025 Labor’s primary is running about 6.7 points under its election vote — a standard post-election fade, not a collapse, on this record.
• Two-party preferred, −3.4 on 2025 The government remains narrowly in front on two-party preferred — about 51.8 versus 55.2 at the election, 3.4 points off its election pace.
• Age of the government, 2nd term This is the 2nd consecutive Labor term. More than any single poll, age sorts the fates: no first-term government on this record has fallen, and the six that did were all at least one term in — two in their second, two in their third, one in its fourth and one in its fifth. The Gillard and second Whitlam governments both fell at exactly this age.

The blend is the point: an ordinary post-election fade in the primary, a two-party position still in front, in a young government. The model’s read is unremarkable — the historical company of terms that survived.

The nineteen terms it learned from

Every few months of each completed term, from month six to the final pre-election read, the model says how likely that government was to be ousted at that point. The six defeats are in colour; the current term is pinned underneath. Terms shorter than a checkpoint show a dot where they never lived to see one.

† The 2007 term ran 33 months, so its month-30 read doubles as its final read. Table values are the model’s in-sample reads; the headline call above comes from the bootstrap described below.

The column that matters today is month 18 — the yardstick nearest this term’s age of 17.4 months.

The six eventual losers read 57, 69, 90, 79, 19, and 33 per cent there, the thirteen winners ran from 0 to 71 — and the current term read 30 at that age.

You can score 68 per cent on this table by simply always predicting “returned” — incumbents usually win. Tested honestly, with each term held out in turn and never seen by its judge, the model calls between thirteen and fifteen of the nineteen terms correctly at every checkpoint, the final read included. History blurs at the extremes: the Hawke government’s famous mid-term scare of 1989–90 read as high as 79 per cent ouster-risk at month 15, and it recovered to win the 1990 election — while the 2016–19 Coalition sat at 19 per cent by its final read and survived.

A note on the six oldest terms. The model’s record was extended back to 1974 in September 2026, and the era rows are the roughest on the table: before 1983 no published two-party-preferred series exists, so every era two-party figure is modelled from first preferences and historical preference flows. Era parliaments also ran at irregular lengths — the 1974–75 and 1983–84 terms were over before month 24, and three of the six era terms never saw a three-year read at all. The model takes the record as it finds it; those checkpoints deserve slightly less trust than the modern ones. (The constants and caveats are documented in the modelling suite’s README.)

Where this can go wrong

There are only nineteen terms. Six defeats is still a tiny sample, and even the final-read accuracy of 79 per cent carries roughly seventeen points of statistical noise either way. Treat every number on this page as a best guess with a range attached — which is why the range, not the median, is the most honest figure here.

It cannot see shocks. The model reads the shape of polling, not events. The 2010–13 minority government’s collapse has no antecedent in the record, and the model never calls it — at any horizon. The 2019–22 term flips depending on exactly when you look: the COVID rally peaks at precisely the fifteen-month yardstick. The 1977–80 government is the opposite kind of miss — re-elected comfortably, yet the cross-check model just tips it over fifty-fifty at its live reads. And 1990 is the one genuine comeback on record — a 77 per cent ouster-read at month 12 that still won.

Waiting buys less than you’d think. Accuracy wanders in a 68-to-79-per-cent band from month six onward and no longer climbs at the finish — the final pre-election read, at 79 per cent, is no better than the year-one reads. Governments that are structurally finished are often visible early — three of the six eventual losers were already rated more likely than not to fall by month six — but the era defeats read murkier, and everything else stays murky until the campaign.

Winning means seats, not the popular vote. Fates are coded by who formed government. Howard’s 1998 win on 49 per cent of the two-party vote counts as a win — the model is about survival, not vote share. And none of this is a forecast of events between now and the election: it is a reading of history’s signatures, and the election itself remains its own fact.

The model, technically

The headline is the best of four constructions tried in the site’s re-election modelling suite (.build/analysis, last revised in the September 2026 feature audit — see below). The ladder, with how each reads Albanese’s second term:

The construction. Each completed term contributes snapshots at ages 6, 12, 15, 18, 24, and 30 months plus a final read three months before its last day — a hundred and thirty-one snapshots in all. Each snapshot’s features are trailing-three-month summaries knowable at that age: the primary-vote and two-party swings against the term’s own election result, the government’s incumbency age in consecutive terms, the fraction of the term elapsed (the current term is assumed to run a full span), and the interaction of primary swing with elapsed fraction — five features in all. Missing values are median-imputed within the training fold; features are standardised within it; the fit is ridge logistic regression with λ = 1.

The validation. Leave-one-term-out: a term’s snapshots are scored only by a model trained on the other eighteen. The baseline “always re-elected” scores 68 per cent. Snapshot-level AUC is 0.75 with Brier score 0.185. The interval comes from a 300-draw cluster bootstrap that resamples whole terms, refits, and re-scores the current term each draw. The live call, verbatim:

The feature audit. Until September 2026 this model also read leadership ratings — net PM approval and the preferred-PM lead. Ablated feature by feature on this same leave-one-term-out harness (run over the 1987-and-later thirteen-term record, before the era extension), neither earned its place: dropping the preferred-PM lead changed no accuracy band at all, and dropping both leadership readings lifted snapshot AUC from 0.77 to 0.84 and cut the Brier score from 0.182 to 0.153, with per-band accuracy never worse. Their fitted coefficients had shrunk to nothing, or landed the wrong sign — Howard was popular mid-term and lost anyway. That matches the published record: preferred-PM scores are weak, skewed predictors of election outcomes (Kevin Bonham, 2020), and the strongest leadership construction on offer — the PM-versus-opposition-leader approval margin (Armarium, 2021) — scored no better here than ignoring leadership ratings entirely. The ratings still tell part of the term’s story, so they stay in the live summary as context; the model just no longer lets them vote.

The cross-check. A deliberately different construction — one ridge logistic per term on first-sixteen-month features — reads p(ousted) = 0.27 — same verdict, well short of alarm. Its best-calibrated variant, adding leadership-spill and minority-government flags, reaches 74 per cent leave-one-term-out accuracy with AUC 0.81 and Brier 0.153; adding election-quarter unemployment lands in the same place. On nineteen terms the estimator is not the constraint — a diagonal LDA nearly ties the ridge and k-nearest-neighbours collapses — and no capacity beyond logistic earns its keep: adjacent accuracies are statistically indistinguishable (the base model’s 68 per cent carries a 95% Wilson interval of roughly [46%, 85%]).

To reproduce: from the repo root, node .build/analysis/reelect-snapshot-hazard.mjs (the headline — its snapshot age defaults to the canonical 16.2 months and moves via --age=N.N), node .build/analysis/reelect-term-ridge.mjs (the cross-check), plus reelect-15mo-levels.mjs and reelect-15mo-declines.mjs (the composites). Both models emit machine-readable results with --json; this page is regenerated from them by .build/refresh-prediction.mjs on a daily due gate, and each refresh is one dated, selectable record above. The analysis scripts read poll data straight from origin/main, so they are immune to working-tree state; the canonical numbers live in .build/analysis/README.md. The analysis’s own closing caution stands: this is historical signature analysis, not a forecast.