Opposition
55average seats
LABGRNTPM
48–6390% seat range
10%chance of a seat majority
ELECTION-DAY OUTLOOK
Election-day forecast.
Explore possible outcomes, with the uncertainty in view.
2026 ELECTION OUTLOOK · 7 Nov 2026
Experimental forecastThe race for a majority.
Explore the combinations that could have the seats to govern.
Coalition seat outlook
61 seats for a majority121-seat reference Parliament
Opportunity
7average seats
OPP
0–1190% seat range
Shown separately
Government parties
59average seats
ACTNZFNAT
52–6790% seat range
34%chance of a seat majority
Majority mark: 61Band: 90% rangeShared seat scale: 0–70
Seat totals are simulation averages, rounded to whole seats. Each simulation uses its own Parliament size and majority target. These groups do not imply a coalition agreement.
Seat odds depend on the selected electorate assumption. This experimental model has limited historical evidence; its probability calibration is still being tested.
Latest saved forecastPolls through 10 Oct 2026Checked 12 Oct, 7:05 pm NZDT
Add Opportunity to either combination
Seat arithmetic only. Neither combination implies a political agreement.
National + ACT + NZ First + Opportunity87%majority-seat probability
65.9 mean seats · 58–73 in the 90% range
Labour + Green + Te Pāti Māori + Opportunity61%majority-seat probability
61.9 mean seats · 54–69 in the 90% range
What this seat scenario assumes
The default uses three Te Pāti Māori electorate wins, informed by its leads in the verified Te Tai Hauāuru, Hauraki-Waikato and Waiariki polls. Other electorates remain uncertain; this is a starting assumption, not a probability-ranked forecast. Electorate evidence & limits →
Assumes ACT wins 1, Te Pāti Māori 3, Opportunity 0, Te Tai Tokerau Party 0 and independent Tākuta Ferris 0 electorates. TPM has 1 additional list candidate available. These are assumptions, not electorate forecasts.
Parliament’s 90% range: 120–122 elected MPs. Majority target: 61–62. List shortages leave 0.0 seats unfilled on average. Overhang probability: 75%.
Nomination evidence & electorate limits
Official nominations closed on 9 October. TPM has one list candidate; TTTP has two. Dual-listed electorate winners are removed from available lists. Other parties’ list capacities use conservative overlap assumptions. No probabilities are assigned to these scenarios.
- Confirmed nomination and list counts
- Official party lists
- Electorate boundary changes
- MMP rules
- Electoral Act sections 191–193
Opposition parties are grouped for comparison, without assuming a coalition agreement. Government grouping source. Explore your own scenario →
Party vote & seats Compare current support and election day
The election-day picture.
Allowing for change before election day
28.2%Election-day party vote
22.8–33.7% 90% range
57% chance of the most party votes27.4%Election-day party vote
22.2–32.9% 90% range
43% chance of the most party votesColoured band: 90% range · marker: mean · vote scale: 0–40%
National leads Labour by 0.8 points on average.
The 90% range for that lead is -8.4 to +9.9 points. Either party could finish ahead.
Party vote & seats
Inspect the polls →Vote ranges include future opinion change and historically trained shared polling error. Seats always refer to the election-day forecast.
| Party | VoteMean / 90% range | SeatsMean / 90% range |
|---|---|---|
| National+0.0 pp | 28.2%22.8–33.7% | 34.728–42 |
| Labour+0.1 pp | 27.4%22.2–32.9% | 33.727–41 |
| Green+0.2 pp | 15.0%11.1–19.1% | 18.414–24 |
| ACT+0.2 pp | 9.7%6.7–13.1% | 12.08–16 |
| NZ First-0.2 pp | 10.2%7.2–13.6% | 12.59–17 |
| Te Pāti Māori-0.1 pp | 1.6%0.6–3.0% | 3.13–4 |
| Opportunity-0.1 pp | 6.4%4.0–9.1% | 6.80–11 |
| Other-0.1 pp | 1.6%0.5–3.0% | 0.00–0 |
Party means sum to 100% before rounding. Individual intervals and medians do not add up. “Other” comprises separate ineligible parties and receives no combined seats.
Snapshot details & downloads
Published 10 Oct, 9:59 am NZDT · nz-polling-bayesian-v14Movement since 8 Oct 2026Forecast detailMethod & limitsSnapshot data
Exact 6-poll prior refit passed for this snapshot · largest mean change 0.13 pp. Review the checks and five-cycle comparison →
Validation, benchmark & source evidence
The underlying model passed convergence checks. Its joint uncertainty was checked using information available at historical cutoffs, without using the target election result to train its error allowance. There are 37 verified historical cutoffs; the primary comparison covers 4 complete election cycles. Checks within an election are correlated.
94.2%historical coverage of 90% party-vote ranges0.986Bayesian CRPS1.010recency benchmark CRPS
Lower CRPS is better. The aggregate advantage is concentrated in 2020; the simpler benchmark does better in the other three complete cycles and at 21, 30 and 31 days. This limited record does not establish a universal model advantage or calibrated coalition probabilities. Historical seat/event checks are conditional on supplied electorate results.
Compare the recency benchmark
| Party | Bayesian | Recency |
|---|---|---|
| National | 28.2%22.8–33.7% | 28.4%23.9–32.9% |
| Labour | 27.4%22.2–32.9% | 28.0%23.6–32.4% |
| Green | 15.0%11.1–19.1% | 15.1%11.7–18.8% |
| ACT | 9.7%6.7–13.1% | 9.7%6.9–12.8% |
| NZ First | 10.2%7.2–13.6% | 10.4%7.6–13.6% |
| Te Pāti Māori | 1.6%0.6–3.0% | 1.5%0.5–2.8% |
| Opportunity | 6.4%4.0–9.1% | 5.9%3.8–8.4% |
| Other | 1.6%0.5–3.0% | 1.1%0.3–2.3% |
Thresholds, diagnostics & sensitivity
- National: >99% model probability of at least 5% party vote. Electorate qualification is separate.
- Labour: >99% model probability of at least 5% party vote. Electorate qualification is separate.
- Green: >99% model probability of at least 5% party vote. Electorate qualification is separate.
- ACT: >99% model probability of at least 5% party vote. Electorate qualification is separate.
- NZ First: >99% model probability of at least 5% party vote. Electorate qualification is separate.
- Te Pāti Māori: <1% model probability of at least 5% party vote. Electorate qualification is separate.
- Opportunity: 81% model probability of at least 5% party vote. Electorate qualification is separate.
All sampled variables: maximum R-hat 1.0034, minimum bulk ESS 1688, minimum tail ESS 3070, 0 divergent transitions. 8,000 joint simulations. Posterior predictive checks are in-sample checks, not election calibration.
5 of 6 conditional prior-sensitivity checks have adequate importance ESS. These checks cannot establish robustness outside the stored posterior.
The doubled house-effect prior has now been refitted exactly on this input: maximum mean change 0.13 pp and event change 0.82 pp across 16 electorate assumptions. The saved low-ESS importance result remains archived; the exact refit supplies the sensitivity evidence.
- Four complete historical elections; repeated parties and cutoffs are dependent.
- Bayesian CRPS is lower overall, but the advantage is concentrated in 2020; recency scores better in the other three complete cycles and at 21, 30 and 31 days.
- Observed 90% coverage is descriptive, not certification of future event probabilities.
- Seats and mathematical majorities are conditional on declared electorate assumptions; no government-formation probabilities.
- Mode, population, fundamentals and pollster-cycle effects are not estimated; sparse current data cannot identify them reliably.
- The 28 September Verian poll is included at revision 2, using the verified original party breakdown. This corrects a stale exclusion note in the saved metadata; the recorded model input and numerical snapshot are unchanged.
Read the explicit metadata correction. Canonical snapshot bytes retain their original hash.
Polls used & excluded
- Curia Market Research · published 8 Oct 2026 — included · revision 1
- Roy Morgan · published 31 Aug 2026 — included · revision 1
- Roy Morgan · published 6 Oct 2026 — included · revision 1
- Reid Research · published 6 Oct 2026 — included · revision 1
- Verian · published 28 Sept 2026 — included · revision 2
- Verian · published 6 Oct 2026 — included · revision 1
Undecided/non-response figures and the respondent-base transformation remain visible in the poll database.
Inference completed 10 Oct, 9:47 am NZDT; inference was completed offline. Input version 59a4c9a350176718 · snapshot 39d6ffbc5dca2f47. Canonical snapshot bytes reproduce that SHA-256 hash. Numerical assumptions and the prospective validation record are versioned separately from these live snapshots.
Forecast checks & research evidenceHistorical comparisons, source recovery and prospective testing
PREDICTION DEVELOPMENT · V20
What the new evidence changes.
Two complete campaign-period Roy Morgan releases have been recovered and tested in six exact historical fits. A third recovered release remains excluded because a minor-party value is censored. The prospective register is still empty: its frozen prediction awaits surveys fielded from 10 October onwards. The live model and priors are retained.
0.13 pplargest mean change in the exact six-poll refit5 cycles40 dependent historical cutoffs13 pairsverified historical electorate observations
Campaign coverage recovered: two complete original releases
Two Roy Morgan releases now fill gaps in the 2011 and 2014 campaign archives. The 31 October 2011 release covers 10–23 October, with 933 electors and 6.5% not naming a party. The 3 September 2014 release covers 18–31 August, with 762 electors and 3.5% not naming a party. Both retain the original complete rounded party vector; their decided sample bases are approximations.
| Election / days left | Before addition | With addition | Recency with addition |
|---|---|---|---|
| 2011 / 1 | 1.112 | 1.081 | 1.088 |
| 2011 / 7 | 1.637 | 1.657 | 1.573 |
| 2011 / 14 | 1.643 | 1.660 | 1.573 |
| 2014 / 1 | 0.610 | 0.673 | 0.649 |
| 2014 / 7 | 0.736 | 0.797 | 0.803 |
| 2014 / 14 | 0.739 | 0.795 | 0.806 |
2011: average Bayesian CRPS rose from 1.464 to 1.466. Coverage of 90% intervals moved from 76.2% to 76.2%.
2014: average Bayesian CRPS rose from 0.695 to 0.755. Coverage of 90% intervals moved from 100.0% to 100.0%.
Six exact fits passed sampling checks, and every earlier control reproduced its saved joint draws and scores. Model settings and original uncertainty-training anchors stayed fixed. These are two elections with three dependent cutoffs each. Both additions are from the same pollster; campaign coverage from other organisations remains incomplete.
The 19 November 2011 release is inspectable but excluded: United Future is censored below 0.5% in the table, with a conflicting sign in the narrative. Missing shares are not inferred as zero. The two earlier Newshub recoveries also remain excluded.
Historical recovery does not change the frozen prospective prediction or establish calibration. Qualifying new surveys must have fieldwork wholly within 10 October–6 November 2026.
2011 original release ↗ · 2014 original release ↗ · Full matched scores · Source audit · Method
How much does the prediction depend on each pollster?
We refitted the exact six-poll Bayesian snapshot four times, removing every survey from one organisation at a time. This tests source dependence. It does not tell us which pollster is right or measure election-day accuracy.
| Pollster removed | Surveys | Current support | Election-day mean | Event probability |
|---|---|---|---|---|
| Curia | 1 | 0.57 pp | 0.59 pp | 9.2 points |
| Reid Research | 1 | 0.89 pp | 0.86 pp | 10.2 points |
| Roy Morgan | 2 | 1.11 pp | 1.10 pp | 14.7 points |
| Verian | 2 | 0.43 pp | 0.41 pp | 6.6 points |
Removing Curia, Reid Research, Roy Morgan, Verian crossed at least one disclosure threshold: a 0.5-point party-mean change or a 5-point event-probability change. Event changes include party leadership, the 5% party threshold and majority-seat combinations under 16 separate electorate scenarios.
Changing the source set also changes sample size, fieldwork coverage and the model’s centred pollster effects. These results are influence checks, not a reason to discard an inconvenient poll. Every refit uses the same priors and historical election-error allowance. The unchanged reference was reproduced from its saved joint draws; all refits passed convergence checks.
The live prediction and frozen prospective test are retained. This review applies to the dated snapshot; later inputs require a fresh study.
Download full party and conditional seat comparisons · Read the method
New original report: two pollsters in the 2011 archive
The V18 archive snapshot recovered the original May 2011 Reid Research report and contained four polls from two pollsters. That survey interviewed 1,000 eligible voters on 20–26 May, with a reported decided-party base of 871. It preserves 11.1% undecided and 1.8% who would not vote. Later campaign additions are compared separately above.
The second pollster’s observation is 154–183 days old at these campaign cutoffs. It helps extend the archive; fresh campaign coverage from other pollsters remains a gap.
| Days to election | Original Bayesian | Expanded archive | Recency |
|---|---|---|---|
| 1 | 1.083 | 1.112 | 1.116 |
| 7 | 1.461 | 1.637 | 1.498 |
| 14 | 1.470 | 1.643 | 1.498 |
| 30 | 1.812 | 2.074 | 2.041 |
| Cutoff average | 1.457 | 1.617 | 1.538 |
The average Bayesian error score rose from 1.457 to 1.617. Its 90% interval coverage was 78.6% before the addition and 75.0% afterwards. The original control was reproduced from the exact saved draws and scores.
The model, priors, uncertainty training and benchmark windows stayed fixed. Four cutoffs from one election are dependent evidence, and past-only 2011 has no earlier election to train its miss allowance. This experiment does not select a new live model or establish probability calibration. Retrospective leave-one-election-out scores and conditional seat errors are recorded separately in the audit.
Original report ↗ · Inspect the archive · Download matched scores · Source and availability rules
Larger common polling misses: a controlled five-election test
We tested a heavier-tailed election-wide error distribution using the same 40 saved Bayesian polling fits. The polling model still uses PyMC; this comparison changes only the final joint error layer. Seven degrees of freedom were fixed before scoring, rather than fitted to five election outcomes.
| Model | CRPS | 90% coverage | Threshold Brier |
|---|---|---|---|
| Live Bayesian rule | 1.098 | 90.4% | 0.040 |
| Same-anchor Dirichlet control | 1.105 | 92.1% | 0.041 |
| Heavier-tail challenger | 1.098 | 92.5% | 0.041 |
| Recency benchmark | 1.139 | 85.4% | 0.033 |
The challenger improved CRPS against both Bayesian controls in 2 of five cycles. It did not pass all the declared improvement checks, so it has not replaced the live prediction. An additional random seed checks whether simulation noise changes the conclusion.
Each election contributes one historical error observation at the selected anchor. The eight cutoffs within an election are averaged before combining elections. Future results and the entire target election are excluded from past-only training. These are retrospective model-development results; they cannot establish probability calibration.
The normal comparison had undefined seat scenarios at 1 of the 80 dependent test rows; the Student-t challenger had 3. All 8,000 vote draws at each cutoff remain in the vote scores. An affected scenario’s seat score is withheld; failed draws are not removed or reweighted. This follow-up was separately versioned after the first seat calculation stopped.
On the 10 Oct 2026 six-poll input, the prior stress test covered all 16 conditional seat scenarios. Changing the challenger’s error-prior dispersion moved party means by at most 0.02 percentage points and event probabilities by 5.05 percentage points. This did not meet the declared robustness limits. It tests this research distribution, not the retained live model. Later inputs require their own checks.
Read the specification → · Download the full comparison →Exact six-poll refit: the prior passed this robustness check
Doubling the pollster-effect prior on the 10 Oct 2026 six-poll input moved party means by at most 0.13 percentage points and event probabilities by 0.82 percentage points across the 16 nomination-aware seat scenarios. Both stayed below the predeclared limits of 0.5 and 5 points. Maximum R-hat was 1.0030, with no divergences.
The unchanged forecast was reproduced from its exact joint draws before comparison. This replaces the unreliable doubled-house-prior importance check for that snapshot. It is a same-input sensitivity result; it does not establish forecast calibration. Later inputs require their own checks.
Five elections: Bayesian and recency compared honestly
We tested the added historical uncertainty using 20 anchors and 40 already verified polling fits, at 1, 3, 7, 10, 14, 21, 30 and 31 days before each election. Past-only training excludes the target election and all results unavailable at its cutoff. Leave-one-election-out results are reported separately. The eight cutoffs within each cycle are dependent.
| Election | Bayesian | Recency |
|---|---|---|
| 2011 | 1.544 | 1.654 |
| 2014 | 0.748 | 0.720 |
| 2017 | 1.033 | 1.032 |
| 2020 | 1.019 | 1.243 |
| 2023 | 1.144 | 1.046 |
| Equal-cycle average | 1.098 | 1.139 |
Recency has the lower error score in three of five cycles and the better threshold Brier score overall. The Bayesian average advantage is concentrated in 2011 and 2020. Its 2011 coverage of 90% intervals remains 75.0%.
The expanded-error challenger raised 90% coverage from 90.4% to 92.1%, but worsened CRPS from 1.098 to 1.105 and worsened threshold Brier score. It failed the declared improvement rule, so we retained the live error model. Pollster coverage and MAE, RMSE, log, energy, threshold and conditional seat scores are in the audit.
The 2011 sources remain inspectable in the poll database. Archive expansion followed earlier diagnosis; five cycles cannot certify general calibration.
Download source audits and benchmark evidence →Joint electorate model: fitted, tested, still in research
We recovered 15 original candidate-poll/result pairs across 3 elections. Two 2017 records lack a verified candidate-response denominator and are excluded. The 13 eligible pairs come from 2 cycles, all from Curia; they are not 13 independent elections.
The PyMC model partially pools the TPM-versus-Labour candidate-pair error and allows a common election-wide miss. It preserves the published undecided and unreported responses. A probability of finishing above Labour is a pairwise endpoint; other candidates can still win the electorate.
| Error model | CRPS ↓ | Brier ↓ |
|---|---|---|
| Sampling only | 7.54 | 0.388 |
| Pooled extra error | 4.72 | 0.158 |
| Shared election error | 5.29 | 0.179 |
Extra error helped compared with textbook sampling alone. The shared-error version did not improve the simpler comparator’s held-out scores. All sampled parameters passed convergence checks, and an exact doubled-shared-prior refit is archived. The small cycle count, changed boundaries, competitive third candidates and missing link to national vote keep unconditional electorate and seat probabilities disabled. The live nomination-aware seat selector continues to show separate, unweighted assumptions.
Read the technical specification and publication limits →Inspect the four current electorate polls
These surveys cover local enrolled Māori voters. Percentages include undecided respondents; they are not national party support.
Te Tai Tonga
Curia for Whakaata Māori · 14 Sept 2026–24 Sept 2026 · n=500
- Mananui Ramsden
- 30%
- Lisa Marie Murch
- 17%
- Lisa Te Morenga
- 16%
- Tākuta Ferris
- 15%
- Undecided
- 17%
Weighted by age and gender for the Māori adult population. Named candidates and undecided total 95%; the remaining responses are unreported and are not redistributed.
Original commissioner release ↗Te Tai Hauāuru
Curia for Whakaata Māori · 14 Sept 2026–24 Sept 2026 · n=500
- Debbie Ngarewa-Packer
- 38%
- Te Pūoho Kātene
- 27%
- Coral Raukawa
- 10%
- Undecided
- 18%
- Other responses
- 6%
Census gender targets; age targets use 2023 electorate turnout. Rounded all-respondent totals 99%. Published decided percentages are retained separately: 46%, 34%, 13%, Other 7%.
Original commissioner release ↗Hauraki-Waikato
Curia for Whakaata Māori · 21 Sept 2026–1 Oct 2026 · n=500
- Hana-Rawhiti Maipi-Clarke
- 45%
- Kingi Kiriona
- 26%
- Undecided
- 17%
- Other responses
- 13%
Census gender targets; age targets use 2023 electorate turnout. Rounded all-respondent totals 101%. Published decided percentages: 54%, 31%, Other 15%.
Original commissioner release ↗Waiariki
Curia for Whakaata Māori · 21 Sept 2026–1 Oct 2026 · n=500
- Rawiri Waititi
- 42%
- Tania Waikato
- 16%
- Toni Boynton
- 14%
- Pāpā Wharewera
- 3%
- Undecided
- 18%
- Other responses
- 6%
Census gender targets; age targets use 2023 electorate turnout. Rounded all-respondent totals 99%. Published decided percentages: 52%, 20%, 18%, 3%, Other 7%.
Original commissioner release ↗Prospective testing stays frozen
At the 10 Oct 2026 check, 0 surveys qualified. Eligible fieldwork must fall wholly within 10 Oct 2026–6 Nov 2026. The original scoring workflow checked nine candidates: five starting inputs, one earlier-fieldwork survey and three historical recoveries were excluded. Inspect each eligibility decision.
Scheduled checks score new eligible surveys without refitting or replacing the registered prediction. This is a dated record; see the current survey count. An empty cohort has no observed accuracy or calibration result.
Archived research & earlier validation reports
These reports are preserved as written. Their publication-gate statements describe the release state when each report was produced. The current experimental release, its limits and the simpler benchmark appear above.
FULL-PARTY UNCERTAINTY · RESEARCH V4
Testing prediction closer to election day.
Historical predictions now cover 1, 7, 14 and 30 days before election day. Each uses the polls available then. Election error is learned separately at each horizon, with the scored election excluded.
Full-party CRPS · lower is better1.021Trained recency benchmark 1.011 · lower error in 2/5 elections
Observed 90% interval coverage91.3%Trained benchmark 88.3% · 5 independent elections
Mean interval width: 6.62 percentage points; trained benchmark 5.09. Mean vote-share error: 1.44 points. Each election counts once within this horizon; all party categories, including Other, are scored. A common-party comparison is also retained.
2011 · 2011-11-12 · 1 polls / 1 pollsters
Bayesian model + matched error1.00889% interval coverage
Trained recency benchmark1.10089% interval coverage
Newest fieldwork ended 6 days before the cutoff.
Misses, alternatives and training
Outside the model’s 90% intervals: Other. Joint energy score: 3.953 model / 4.433 trained benchmark.
CRPS with election-eve error instead: 1.008. Original fixed-error Bayesian CRPS: 1.005. Common-party CRPS: 1.147 / 1.350.
Training elections: none — declared prior only. Each contributes one vector at 14 days. Only results released by this cutoff enter training.
2014 · 2014-09-06 · 2 polls / 1 pollsters
Bayesian model + matched error0.92580% interval coverage
Trained recency benchmark0.86390% interval coverage
Newest fieldwork ended 20 days before the cutoff. Sparse coverage limits this test.
Misses, alternatives and training
Outside the model’s 90% intervals: NZ First, Conservative. Joint energy score: 3.743 model / 3.600 trained benchmark.
CRPS with election-eve error instead: 0.928. Original fixed-error Bayesian CRPS: 0.971. Common-party CRPS: 0.890 / 0.821.
Training elections: 2011. Each contributes one vector at 14 days. Only results released by this cutoff enter training.
2017 · 2017-09-09 · 1 polls / 1 pollsters
Bayesian model + matched error1.311100% interval coverage
Trained recency benchmark1.29488% interval coverage
Newest fieldwork ended 27 days before the cutoff. Sparse coverage limits this test.
Misses, alternatives and training
Outside the model’s 90% intervals: none in this test. Joint energy score: 5.403 model / 5.138 trained benchmark.
CRPS with election-eve error instead: 1.324. Original fixed-error Bayesian CRPS: 1.327. Common-party CRPS: 1.479 / 1.461.
Training elections: 2011, 2014. Each contributes one vector at 14 days. Only results released by this cutoff enter training.
2020 · 2020-10-03 · 4 polls / 2 pollsters
Bayesian model + matched error0.726100% interval coverage
Trained recency benchmark0.88088% interval coverage
Newest fieldwork ended 12 days before the cutoff.
Misses, alternatives and training
Outside the model’s 90% intervals: none in this test. Joint energy score: 2.809 model / 3.680 trained benchmark.
CRPS with election-eve error instead: 0.701. Original fixed-error Bayesian CRPS: 0.705. Common-party CRPS: 0.814 / 0.995.
Training elections: 2011, 2014, 2017. Each contributes one vector at 14 days. Only results released by this cutoff enter training.
2023 · 2023-09-30 · 9 polls / 2 pollsters
Bayesian model + matched error1.13788% interval coverage
Trained recency benchmark0.91688% interval coverage
Newest fieldwork ended 4 days before the cutoff.
Misses, alternatives and training
Outside the model’s 90% intervals: ACT. Joint energy score: 4.528 model / 3.414 trained benchmark.
CRPS with election-eve error instead: 1.139. Original fixed-error Bayesian CRPS: 1.125. Common-party CRPS: 1.282 / 0.986.
Training elections: 2011, 2014, 2017, 2020. Each contributes one vector at 14 days. Only results released by this cutoff enter training.
19 verified fits span 5 independent elections. Repeated horizons do not add independent elections. This archive cannot establish that a stated 70% probability occurs 70% of the time. Model design followed inspection of the archive; prospective confirmation is still needed. Public forecast publication remains closed.
Audit the method and retained evidence
A mean-preserving Dirichlet discrepancy replaces the fixed election miss across the full composition. Small parties can have larger relative errors. One joint draw represents a possible election; percentages sum to 100%. Conditional pairwise correlations are negative; latent polling uncertainty contributes other dependence. A flexible empirical correlation matrix is not claimed.
The recency benchmark independently learns the same uncertainty family from the same training elections. Latest-poll, recent-average and fixed-error comparisons, party errors, probability scores, prior sensitivity and conditional seat errors are in the full report. Wider intervals alone are not evidence of a better model.
R-hat ≤ 1.0063; minimum bulk/tail ESS 1340 / 1387; 0 divergences. Quadrature checks passed. Numerical success is separate from forecast calibration.
The v4 review estimates each anchor separately. The v5 comparison below tests how error distributions can be transferred to the dates between them; the earlier reports remain unchanged.
Read the technical method · Download the complete horizon report · Earlier month-ahead experiment · Source recovery audit
Report 1046d705fc96256f9e24d101fba9f4bc2263507569ebef3c96e1d8e5fc2e66bf. Generated 2026-10-07T23:20:10.694864+00:00. Source snapshots and selected joint draws are versioned; earlier reports remain available.
BETWEEN THE HORIZONS · RESEARCH V5
Does smoother uncertainty predict better?
New historical cutoffs test 3, 10, 21 and 31 days before election day. The proposed rule mixes the error distributions from adjacent 1, 7, 14 and 30-day anchors. Each prediction uses a polling fit from its exact cutoff.
Simpler nearest-anchor rule retained for owner research
The mixture did not meet every comparison criterion set before these cutoffs were scored. Public forecast publication remains closed.
Calendar mixture CRPS · lower is better1.049Nearest anchor 1.051 · trained recency 1.011
Observed 90% interval coverage91.9%Nearest anchor 94.4% · 4 independent elections
Mean interval width 6.63 percentage points; nearest anchor 6.59. Election-eve-error CRPS 1.046. Smoothness alone does not establish accuracy or calibration.
2014 · 2014-08-30 · 2 polls / 1 pollsters
Calendar mixture CRPS0.931
Trained recency CRPS0.861
Latest fieldwork ended 13 days before the cutoff.
Training and missed intervals
14-day anchor: 56.3% of the error mixture. Training elections: 2011. One vector per election in this anchor.
30-day anchor: 43.8% of the error mixture. Training elections: none — declared prior only. One vector per election in this anchor.
Outside the mixture’s 90% intervals: nz-first, conservative.
Common-party CRPS: 0.895 / 0.819 trained recency. Joint energy: 3.767 / 3.589.
2017 · 2017-09-02 · 1 polls / 1 pollsters
Calendar mixture CRPS1.310
Trained recency CRPS1.300
Latest fieldwork ended 20 days before the cutoff.
Training and missed intervals
14-day anchor: 56.3% of the error mixture. Training elections: 2011, 2014. One vector per election in this anchor.
30-day anchor: 43.8% of the error mixture. Training elections: 2014. One vector per election in this anchor.
Outside the mixture’s 90% intervals: none in this test.
Common-party CRPS: 1.478 / 1.468 trained recency. Joint energy: 5.419 / 5.150.
2020 · 2020-09-26 · 4 polls / 2 pollsters
Calendar mixture CRPS0.737
Trained recency CRPS0.879
Latest fieldwork ended 5 days before the cutoff.
Training and missed intervals
14-day anchor: 56.3% of the error mixture. Training elections: 2011, 2014, 2017. One vector per election in this anchor.
30-day anchor: 43.8% of the error mixture. Training elections: 2014, 2017. One vector per election in this anchor.
Outside the mixture’s 90% intervals: none in this test.
Common-party CRPS: 0.826 / 0.993 trained recency. Joint energy: 2.816 / 3.667.
2023 · 2023-09-23 · 8 polls / 2 pollsters
Calendar mixture CRPS1.217
Trained recency CRPS1.003
Latest fieldwork ended 4 days before the cutoff.
Training and missed intervals
14-day anchor: 56.3% of the error mixture. Training elections: 2011, 2014, 2017, 2020. One vector per election in this anchor.
30-day anchor: 43.8% of the error mixture. Training elections: 2014, 2017, 2020. One vector per election in this anchor.
Outside the mixture’s 90% intervals: act.
Common-party CRPS: 1.369 / 1.118 trained recency. Joint energy: 4.776 / 3.866.
2011: No complete exact-date model-eligible polls released by cutoff.
Selection, boundaries and reproducibility
18 new cutoff fits reuse 19 verified anchors from 5 independent elections. For selection, horizons are averaged within each election, then elections receive equal weight. The mixture must improve past-only CRPS, preserve observed 90% coverage and improve CRPS in at least three elections.
Equal-election past-only CRPS: mixture 0.990, nearest anchor 0.990, independently trained recency 1.025. Mixture improves over nearest anchor in 2/5 elections.
Anchor distributions are mixed, never multiplied as independent evidence. At 31 days, the rule retains 30-day error and uses the exact extra day of polling movement. The offline updater refuses horizons outside 1–31 days. Total uncertainty need not shrink monotonically when new polls arrive.
These cycles were already inspected at the anchor horizons. New cutoffs are not new independent elections or an untouched model-selection archive. This limited comparison does not certify event-probability frequencies or select a public production model.
All fits passed: R-hat ≤ 1.0063, minimum bulk/tail ESS 1340 / 1387, zero divergences.
Complete v5 report · Frozen experiment plan · Technical explanation
Report fb2ab31838fafc047416ab9afd0d034343a29cb5903b9a13c31b63525c9b6f6a · Plan 6d8f33e58ca40771ac666257b593a31bf11a48363ab11e555dec9d56fe58a627 · Generated 2026-10-07T23:59:22.474481+00:00
LEARNED ELECTION ERROR · RESEARCH
Does the uncertainty survive an election test?
The model now learns additional election error from historical outcomes. It accounts for the uncertainty already in the polling estimate and replaces the original fixed allowance. 9 fits pass sampling checks across 5 elections.
Lower CRPS is better: it rewards an accurate prediction with useful uncertainty. The learned model improves this score against the original model in 3 of 5 tests shown. Score differences are small and do not establish a consistent advantage. Coverage below uses the complete party/category table for each cycle, rather than pooling smaller parties into one group.
CRPS compares the same six continuing parties plus a remaining-parties group across cycles. Displayed coverage and interval widths use the full cycle-specific party table, including Other. Grouping minor parties can hide their individual misses; the full report retains both views.
2011 · 2011-11-25 · 2 polls
Learned model CRPS0.9889% coverage · 4.9 pp average interval
Original model CRPS0.9989% coverage · 4.6 pp average interval
Benchmark and training evidence
Fixed recency benchmark CRPS: 1.08. Recency with its own learned error: 1.10. Joint energy score: 3.40 learned / 3.41 original.
Training elections: 2014, 2017, 2020, 2023. The scored election’s outcome is excluded.
2014 · 2014-09-19 · 3 polls
Learned model CRPS0.6780% coverage · 4.6 pp average interval
Original model CRPS0.6470% coverage · 4.2 pp average interval
Benchmark and training evidence
Fixed recency benchmark CRPS: 0.65. Recency with its own learned error: 0.64. Joint energy score: 2.34 learned / 2.25 original.
Training elections: 2011, 2017, 2020, 2023. The scored election’s outcome is excluded.
2017 · 2017-09-22 · 3 polls
Learned model CRPS0.93100% coverage · 5.8 pp average interval
Original model CRPS0.94100% coverage · 5.4 pp average interval
Benchmark and training evidence
Fixed recency benchmark CRPS: 1.03. Recency with its own learned error: 1.04. Joint energy score: 3.62 learned / 3.65 original.
Training elections: 2011, 2014, 2020, 2023. The scored election’s outcome is excluded.
2020 · 2020-10-16 · 7 polls
Learned model CRPS0.9288% coverage · 4.8 pp average interval
Original model CRPS0.9388% coverage · 4.2 pp average interval
Benchmark and training evidence
Fixed recency benchmark CRPS: 1.04. Recency with its own learned error: 1.06. Joint energy score: 3.60 learned / 3.60 original.
Training elections: 2011, 2014, 2017, 2023. The scored election’s outcome is excluded.
2023 · 2023-10-13 · 12 polls
Learned model CRPS1.00100% coverage · 5.6 pp average interval
Original model CRPS1.00100% coverage · 4.9 pp average interval
Benchmark and training evidence
Fixed recency benchmark CRPS: 0.91. Recency with its own learned error: 0.92. Joint energy score: 3.57 learned / 3.61 original.
Training elections: 2011, 2014, 2017, 2020. The scored election’s outcome is excluded.
Sources, assumptions and numerical checks
Each election contributes one terminal-day training vector. The tests use six continuing parties and a remaining-parties group for comparable scoring. The remaining group cannot qualify for seats as one party; separate-party uncertainty inside it remains a declared prior.
The 2011 extension adds two original Roy Morgan releases and official party-vote counts. It is a sparse research archive, separate from the main poll database. Earlier results are admitted only after their recorded availability boundary. No mean correction, economic fundamentals or unrestricted error covariance is fitted.
Worst R-hat 1.0057; minimum bulk/tail ESS 1405 / 1387. All fits have zero divergences. Variance integration passes grid-resolution and tail checks; prior sensitivity and conditional MMP seat scores are in the full report.
Inspect the 2011 sources and transcriptions · Download the full validation report
Report 3ec5f8908fb785002c0420832b6efb50309694431cc6216756e4a37101983068. Updated 2026-10-07T20:32:54.400815+00:00. Public forecast gate remains closed.
ELECTION-DAY TESTS
Testing votes, seats and joint uncertainty
The evaluation is complete. 20 historical fits; 20 pass sampling checks, across 4 elections. Forecast publication remains closed.
Each prediction uses only polls available at its cutoff. Election results enter afterwards, for scoring. Fixed priors are tested on held-out outcomes; this is not a trained leave-one-election-out model.
| Election / cutoff | Vote MAE | Vote CRPS | Joint energy score | 90% party coverage | Seat MAE |
|---|---|---|---|---|---|
| 2014 2014-09-19 | 0.89 / 0.66 | 0.65 / 0.53 | 2.57 / 2.46 | 70% / 90% | 1.05 / 0.93 |
| 2017 2017-09-22 | 1.31 / 1.48 | 0.82 / 0.90 | 3.64 / 3.99 | 100% / 100% | 1.68 / 1.94 |
| 2020 2020-10-16 | 1.21 / 1.36 | 0.82 / 0.92 | 3.59 / 3.99 | 88% / 88% | 1.53 / 1.73 |
| 2023 2023-10-13 | 1.31 / 1.18 | 0.88 / 0.81 | 3.61 / 3.22 | 100% / 100% | 1.72 / 1.48 |
Historical majority-seat events
These are conditional mathematical combinations, not government-formation predictions. No probabilities are assigned to the electorate scenarios.
| Combination | Probability | Observed majority | Brier score |
|---|---|---|---|
| national + act | 28.4% | No | 0.081 |
| national + act + nz-first | 89.5% | Yes | 0.011 |
| labour + green | 0.9% | No | 0.000 |
| labour + green + tpm | 7.0% | No | 0.005 |
What this does and does not validate
Joint draws preserve a 100% party-vote composition. Energy scores assess the whole vector; CRPS assesses individual parties. Party-vote threshold events and conditional majority-seat events have Brier scores in the downloadable report. Reported coverage is descriptive, not a pass certificate.
Electorate scenarios are fixed stress tests, not recovered contemporary electorate forecasts. The incumbent scenario uses the previous general election for covered parties only. Retirements, boundary changes and parties outside the model remain unresolved. Actual electorate winners never enter prediction. Seat allocation uses the same tested MMP engine as the calculator. The later Port Waikato seat is excluded from the 2023 comparison.
Four elections are too few to establish whether a stated 70% event happens 70% of the time. Parties and horizons are dependent. Industry-wide error remains prior-driven; its sensitivity is reported. The extra-error benchmarks are declared scenarios, not fitted NZ error models. No model or prior was selected from these results.
Joint scoring uses a seeded 2,000-draw subset and seat events use 1,000 draws; their maximum marginal subsampling standard error is about 1.6 percentage points, conditional on the saved posterior ensemble. This does not include residual MCMC error or uncertainty about the model assumptions. Full draws are processed offline and useful summaries are retained.
PREDICTIVE BENCHMARKS · ARCHIVE COVERAGE
Does uncertainty improve prediction?
Compare ten archived held-out poll predictions with three simple benchmark centers and two fixed error scenarios. CRPS rewards a useful predictive distribution, balancing its spread and where the observation falls. Lower scores are better. This checks published polls, not election outcomes.
| Held-out poll | Training ends | Benchmark polls | Bayesian / benchmark CRPS | Bayesian / benchmark width | Bayesian / benchmark coverage |
|---|---|---|---|---|---|
| verian-2023-08-21 | 2023-08-11 | 1 | 1.06 / 1.67 | 6.35 / 7.71 | 7 / 8 / 6 / 8 |
| verian-2023-09-20 | 2023-09-15 | 1 | 0.76 / 1.54 | 6.63 / 7.66 | 8 / 8 / 7 / 8 |
| verian-2023-10-11 | 2023-10-06 | 1 | 0.73 / 0.76 | 5.06 / 7.69 | 7 / 8 / 8 / 8 |
Results vary by target and benchmark assumptions. A low CRPS on selected polls does not certify election-day calibration. No benchmark scale or model is selected from these outcomes.
Benchmark construction and scoring limits
Joint logistic-normal draws preserve a 100% composition. Training and target sampling uncertainty enter separately. Training precision assumes independent sampling and approximately common party composition. The extra-error scenario adds a fixed isotropic log-ratio standard deviation of 0.15 once; it does not shrink with the number of polls. That scale is assumed, not estimated from NZ errors.
These simple benchmarks have no estimated house effects or future movement. Unknown sample sizes are excluded; their centers can differ from earlier point-only benchmarks. Gross target sample size is conditional post-fieldwork metadata. Target party percentages never enter predictive draws.
The archived Bayesian report retains marginal scores and intervals, but not joint draws. These comparisons therefore evaluate marginal CRPS only. They cannot establish whether the joint multiparty error structure is correct.
Scoring reference: Gneiting & Raftery (2007).
Where the archive is thin
This coverage table uses the refreshed archive. The benchmark scores above retain their original input snapshots.
| Election | Archived | Model eligible | Eligible pollsters |
|---|---|---|---|
| 1996 | 0 | 0 | 0 |
| 1999 | 0 | 0 | 0 |
| 2002 | 0 | 0 | 0 |
| 2005 | 0 | 0 | 0 |
| 2008 | 0 | 0 | 0 |
| 2011 | 0 | 0 | 0 |
| 2014 | 5 | 3 | 1 |
| 2017 | 8 | 3 | 2 |
| 2020 | 8 | 7 | 2 |
| 2023 | 14 | 12 | 2 |
Early-source recovery: two records, no new model inputs
Two early 2017 pollster-supplied releases are now inspectable. Their wider minor-party nets do not separate TOP; a total of 100% does not prove the model composition is complete.
- morgan-2017-01-20 · fieldwork 2017-01-03 to 2017-01-16 · sample 868. Factual transcription of text explicitly supplied by Roy Morgan and syndicated by interest.co.nz on 20 January 2017. Landline and mobile interviews. Non-response means did not name a party; party-support respondent count not published here. Other combines the disjoint published United Future, Conservative, Internet and Independent/Other figures. TOP is not separately reported; no zero or residual is imputed. Excluded from modelling because the configured composition requires TOP separately. Original pollster link currently returns 404; retained as a provenance lead, not claimed as retrieved.
- morgan-2017-02-28 · fieldwork 2017-01-30 to 2017-02-12 · sample 852. Factual transcription of text explicitly supplied by Roy Morgan and syndicated by interest.co.nz on 28 February 2017. Fieldwork ended 12 February; the host publication date is used conservatively rather than assuming an earlier release date. Landline and mobile interviews. Non-response means did not name a party; party-support respondent count not supplied. Other combines disjoint named minor-party figures and Independent/Others net. TOP is not separately reported; no zero or residual is imputed. Excluded from modelling until a compatible detailed composition can be verified.
Unresolved source leads
- Roy Morgan 17 February–2 March release: Scoop presents a verification wall; original pollster URL returns 404. Indexed excerpts and secondary tables are discovery leads only. No poll added.
- 1 NEWS Colmar Brunton 11–15 February detailed PDF: Detailed original PDF referenced in the bibliography, but not retrieved. Rounded secondary summaries omit minor-party detail. No poll added.
- Newshub–Reid Research March poll: Secondary summary located; original commissioner report and full methodology not recovered. No poll added.
Checked 2026-10-07T16:52:43.131751Z. Missing categories and respondent counts are not imputed.
1996–2011 have no collected polls yet. The 2014 and 2017 archives contain no eligible polls at the 180-, 90- or 60-day cutoffs. More early-cycle releases and pollster diversity are needed before fitting more complex effects.
2023: inspect cutoffs and record exclusions
Cutoffs use final official election dates retrospectively. They are a coverage audit, not operational forecasts using dates announced at the time.
| Days before election | Cutoff | Known eligible | Within 45 days | Pollsters | Latest fieldwork |
|---|---|---|---|---|---|
| 180 | 2023-04-17 | 1 | 1 | 1 | 2023-03-26 |
| 90 | 2023-07-16 | 3 | 1 | 1 | 2023-06-25 |
| 60 | 2023-08-15 | 5 | 2 | 2 | 2023-07-23 |
| 30 | 2023-09-14 | 7 | 2 | 2 | 2023-08-27 |
| 14 | 2023-09-30 | 9 | 4 | 2 | 2023-09-26 |
| 7 | 2023-10-07 | 10 | 4 | 2 | 2023-10-03 |
| 1 | 2023-10-13 | 12 | 5 | 2 | 2023-10-10 |
- morgan-2023-10-13: Eligible, subject to publication cutoff
- verian-2023-10-11: Eligible, subject to publication cutoff
- verian-2023-10-04: Eligible, subject to publication cutoff
- verian-2023-09-27: Eligible, subject to publication cutoff
- verian-2023-09-20: Eligible, subject to publication cutoff
- morgan-2023-09-05: Eligible, subject to publication cutoff
- verian-2023-08-21: Eligible, subject to publication cutoff
- morgan-2023-08-01: Eligible, subject to publication cutoff
- verian-2023-07-17: Eligible, subject to publication cutoff
- morgan-2023-07-04: Eligible, subject to publication cutoff
- morgan-2023-05-30: Eligible, subject to publication cutoff
- morgan-2023-05-02: Month-only fieldwork; No model-eligible complete composition
- morgan-2023-03-31: Eligible, subject to publication cutoff
- morgan-2023-03-10: Month-only fieldwork; No model-eligible complete composition
Reproducibility and remaining limits
- Benchmark configurations declared before scoring; no scale or model chosen from these outcomes. Fixed extra scale 0.15 is an assumed stress scenario, not a fitted NZ error estimate.
- All benchmark draws are joint compositions; only marginal CRPS, interval coverage and widths are comparable with archived Bayesian summaries. Joint Bayesian draws were not stored; no joint distribution score reconstructed from marginal intervals.
- Latest/45-day average/14-day recency centers exclude unknown sample sizes. Earlier point-only baselines can differ because their missing-sample conventions differ.
- Training precision uses independent sampling and approximately common party composition; extra isotropic log-ratio variance is added once and never shrinks with poll count. No estimated shared misses, house effects or future trend in these benchmarks.
- Sampling-only baseline is intentionally incomplete; neither benchmark family certifies calibrated uncertainty.
- Gross target samples and post-fieldwork design metadata are conditional evaluation assumptions. Party-support base or target percentages do not enter prediction.
- Ten dependent targets in four cycles with different party categories: compare within rows; no pooled calibration count, overall MAE or winning model claim.
- Coverage is of the collected archive, not all published NZ polls. Historical horizon cutoffs use final official dates retrospectively, not operational announced-date forecasting. Missing earlier results/calendar remain explicit.
Run 2026-10-07T16:41:38.897450+00:00. Report hash b26dca3ac99bdfa970345bf219cdf3f134982775902909b16fb859bb5a4afd15.
Forecast publication remains blocked. The public polling baseline is unchanged.
Download archived benchmark scores and original coverage · Refreshed coverage and audit hash · Source-recovery ledgerSURVEY DESIGN · CORRECTED HISTORICAL TESTS
How much do survey assumptions matter?
Four earlier held-out polls now use the corrected v2 prediction calculation. Together with six compatible archived tests, the same method spans four election cycles. Priors are unchanged; this small, dependent sample does not establish calibration or select a production model.
| Held-out poll | Training stops | Training polls | Model MAE | Latest-poll MAE | Mean CRPS | Party coverage in 90% intervals |
|---|---|---|---|---|---|---|
| morgan-2014-08-20 | 2014-08-03 | 1 | 1.04 | 0.70 | 0.65 | 10 / 10 |
| morgan-2014-09-17 | 2014-08-31 | 2 | 1.33 | 1.30 | 0.97 | 8 / 10 |
All eight new fits pass sampling diagnostics. The corrected model has lower point error than the latest-poll baseline in one of the four earlier targets. With the six modern targets, the evidence remains mixed; no configuration is selected from these scores.
Conditional respondent-base checks
These calculations hold the training posterior fixed. Factors 1, 1.5 and 2 are assumed variance multipliers, not measured NZ design effects. A reported party-support base is used only where the original source supplies it. These counts are post-fieldwork metadata; the checks are conditional evaluations, not operational predictions made before a survey.
| Respondent base | Count | Factor | Effective n | Mean interval width | Largest mean shift |
|---|---|---|---|---|---|
| Gross sample | 1006 | 1 | 1006 | 9.63 | 0.052 |
| Gross sample | 1006 | 1.5 | 671 | 9.80 | 0.000 |
| Gross sample | 1006 | 2 | 503 | 9.96 | 0.051 |
| Published party-support base | 840 | 1 | 840 | 9.71 | 0.031 |
| Published party-support base | 840 | 1.5 | 560 | 9.90 | 0.031 |
| Published party-support base | 840 | 2 | 420 | 10.10 | 0.091 |
Full refits on two fixed targets
Four additional fits change the sampling covariance in both training and prediction. Each is compared with its factor-1.5 reference using the same sampler, seed and priors. This bounded check does not establish robustness across every pollster, survey mode or forecast horizon.
| Target | Factor | Largest mean shift | Mean 90% width | Width change |
|---|---|---|---|---|
| morgan-2014-09-17 | 1 | 0.40 | 4.27 | -0.62 |
| morgan-2014-09-17 | 2 | 0.35 | 5.47 | 0.58 |
| colmar-2017-09-14 | 1 | 0.25 | 9.46 | -0.35 |
| colmar-2017-09-14 | 2 | 0.25 | 10.30 | 0.50 |
Limits and audit trail
- Factors 1, 1.5 and 2 are predeclared stress scenarios; no empirical NZ survey design effect is estimated or selected.
- Conditional target sensitivity holds training posterior fixed; full refits change training and target sampling covariance. Neither establishes calibration.
- Reported party-support bases are post-fieldwork respondent counts, used only in conditional target-design scenarios; not available for operational pre-fieldwork forecasts.
- Only two targets receive full refits. No general robustness claim across pollsters, modes or forecast horizons.
- Corrected reference results plus compatible modern v2 results span ten dependent polls in four cycles. Party categories differ; no pooled calibration count or overall MAE.
- Shared polling misses, electorate assumptions and election-day probability calibration remain unvalidated.
Run 2026-10-07T16:32:26.820304+00:00. Report hash 998c5df5439d1427e7616cd96dc95f90020a110bb7db4ff9c605ba0ef748c7e7. All planned runs recorded.
Disclosure guidance: AAPOR · Pew survey methodology. These references explain survey-error treatment; they do not validate the chosen stress factors for NZ polls.
The public polling baseline remains unchanged. Election probabilities remain withheld.
Download corrected reference fits, conditional checks and full-refit diagnosticsNEW SOURCES · 2020 & 2023
Broader evidence, unchanged priors
Three additional original reports fill gaps before the final campaign polls. Their detailed tables and respondent bases are preserved. A broader archive improves the opportunity to test a model; it does not guarantee a better forecast.
1 NEWS Colmar Brunton Poll · 2020-09-22
Fieldwork 2020-09-17 to 2020-09-21. Gross sample 1008; party-support base 847.
Original preliminary report. Detailed party table and cover show 17–21 September; methodology prose says Monday 22, an inconsistent day/date. Preserve 21 from cover/table and disclose discrepancy. 405 landline and 603 mobile interviews; weighted age, gender, region, ethnicity and phone access. Party-support base 847; published undecided/refused total 14%. Other sums minor categories once; dash entries contribute no separate positive share. Rounded table totals 99.3%.
Original detailed report1 NEWS Verian Poll · 2023-08-21
Fieldwork 2023-08-12 to 2023-08-16. Gross sample 1002; party-support base 792.
Commissioner-uploaded original report, detailed party table. 501 mobile and 501 online interviews; weighted age, gender, region, ethnicity and education. Probed party-support base 792; undecided/refused total 12%. Other includes Freedoms NZ net once; its component rows are not added again. Detailed NZ First 3.7%, TPM 2.6% and TOP 0.6% take precedence over rounded summary. Total 100.2% retained.
Original detailed report1 NEWS Verian Poll · 2023-09-20
Fieldwork 2023-09-16 to 2023-09-19. Gross sample 1001; party-support base 807.
Commissioner-uploaded original report, detailed party table. 501 mobile and 500 online interviews; weighted age, gender, region, ethnicity and education. Probed party-support base 807; undecided/refused total 12%. Other counts Freedoms NZ net once, excluding its repeated component rows. Net 0.7% differs from rounded component sum 0.8%; preserve net. Total 100.7% retained.
Original detailed reportSix pre-fieldwork prediction tests
These 2020 and 2023 targets were fixed before fitting. Only polls available before target fieldwork began enter inference. Priors and the seed remain unchanged. Target results and election outcomes enter neither fitting nor prediction; source design metadata is used conditionally for evaluation.
| Held-out poll | Training stops | Training polls | Model MAE | Latest-poll MAE | Mean CRPS | Party coverage in 90% intervals |
|---|---|---|---|---|---|---|
| colmar-brunton-2020-09-22 | 2020-09-16 | 3 | 1.48 | 1.39 | 1.02 | 8 / 8 |
| colmar-brunton-2020-10-08 | 2020-10-02 | 4 | 1.15 | 0.57 | 0.78 | 7 / 8 |
| colmar-brunton-2020-10-15 | 2020-10-09 | 5 | 0.57 | 0.67 | 0.48 | 8 / 8 |
| verian-2023-08-21 | 2023-08-11 | 5 | 1.53 | 2.34 | 1.06 | 7 / 8 |
| verian-2023-09-20 | 2023-09-15 | 7 | 1.20 | 2.23 | 0.76 | 8 / 8 |
| verian-2023-10-11 | 2023-10-06 | 10 | 1.09 | 0.84 | 0.73 | 7 / 8 |
6 of six fits pass sampling diagnostics. Among passing fits with a latest-poll comparison, the model has lower point error in 3 of 6 tests. Compare within rows; do not pool the dependent parties into a calibration sample size.
Predictive method revision and remaining limits
Predictive log-ratio mean now matches the fitted likelihood directly, without applying a second continuity adjustment. Sampling covariance uses expected reported support including house effects. Old v1 reports remain preserved and are not pooled with these scores.
This corrects the predictive calculation, not the model priors. Previous reports remain archived and are not pooled with these v2 scores.
- Six dependent target polls across two additional cycles. No election-day calibration certification.
- Training histories overlap; party intervals are correlated. Do not pool party coverage into an independent sample size.
- Only two pollster lineages represented. Selected archive remains incomplete.
- Gross sample proxy can understate sampling error; design-effect and decided-base sensitivity remain outstanding.
- The 2020 preliminary source contains a methodology day/date typo; cover and party table agree on 17–21 September and those dates are used.
- Predictive method v2 corrects the log-ratio mean/covariance alignment. Earlier v1 held-out scores are historical snapshots and are not combined.
- No prior selection or industry-miss calibration from these outcomes.
Run 2026-10-07T09:39:09.253025+00:00. Input hash e10115d5d24b15d7ac5369b7db5f00b96238cfc6860adad27c57b24546b7e9aa.
Updated shared-error audit
The source expansion also updates the retrospective leave-one-election-out point-error comparison. Held-out election outcomes are excluded from its training mean. This audit does not calibrate shared polling uncertainty.
| Held-out election | Uncorrected | Training-mean correction |
|---|---|---|
| 2023 | 1.56 | 1.48 |
| 2020 | 1.45 | 2.23 |
| 2017 | 1.92 | 1.96 |
| 2014 | 0.85 | 0.78 |
Four independent cycles remain insufficient to establish election probability calibration. The public baseline stays available; forecast probabilities remain withheld.
Held-out predictions, party intervals and diagnostics · Updated point baselines · Updated shared-error audit
SOURCE EXPANSION · UPDATED RESEARCH
Original reports recovered
The archive now includes original September 2014 Roy Morgan and 3 News–Reid reports, plus the detailed September 2017 Colmar Brunton PDF. Raw figures, sample bases and source links are preserved. Earlier research below remains a dated snapshot.
Roy Morgan · 2014-09-17
Fieldwork 2014-09-01 to 2014-09-14. Sample 935.
Complete composition · eligible for offline research.
Original archived reportReid Research · 2014-09-10
Fieldwork 2014-09-02 to 2014-09-08. Sample not stated.
Denominator unverified · excluded from all modelling.
Original archived reportReid Research · 2014-09-17
Fieldwork 2014-09-09 to 2014-09-15. Sample 1000.
Denominator unverified · excluded from all modelling.
Original archived report1 News Colmar Brunton Poll · 2017-09-14
Fieldwork 2017-09-09 to 2017-09-13. Sample 1006; party-support base 840.
Complete composition · eligible for offline research.
Original archived reportThe 2017 PDF’s detailed table reports ACT at 0.6% and Opportunity at 1.6%; its summary rounds these to 1% and 2%. We use the detailed table and its reported party-support sample of 840. The published combined undecided figure is 14%; rounded components must not be added to replace it.
What changed in the backtests?
| Election / days | Pollsters | Latest fieldwork age at cutoff | Bayesian | Recency baseline | Sampling |
|---|---|---|---|---|---|
| 2014 / 180 | 0 | — | Withheld | — | No complete exact-date model-eligible polls released by cutoff. |
| 2014 / 90 | 0 | — | Withheld | — | No complete exact-date model-eligible polls released by cutoff. |
| 2014 / 60 | 0 | — | Withheld | — | No complete exact-date model-eligible polls released by cutoff. |
| 2014 / 30 | 1 | 4 days | 1.24 | 1.18 | Passed |
| 2014 / 14 | 1 | 20 days | 1.27 | 1.18 | Passed |
| 2014 / 7 | 1 | 27 days | 1.26 | 1.13 | Passed |
| 2014 / 1 | 1 | 5 days | 0.88 | 0.59 | Passed |
| 2017 / 180 | 0 | — | Withheld | — | No complete exact-date model-eligible polls released by cutoff. |
| 2017 / 90 | 0 | — | Withheld | — | No complete exact-date model-eligible polls released by cutoff. |
| 2017 / 60 | 0 | — | Withheld | — | No complete exact-date model-eligible polls released by cutoff. |
| 2017 / 30 | 1 | 11 days | 2.08 | 1.84 | Passed |
| 2017 / 14 | 1 | 27 days | 2.07 | 1.84 | Passed |
| 2017 / 7 | 2 | 3 days | 1.31 | 1.49 | Passed |
| 2017 / 1 | 2 | 9 days | 1.28 | 1.49 | Passed |
The first expanded 2014 final-cutoff run narrowly failed the R-hat threshold (1.0112 against 1.01). The table uses a separately recorded longer run with unchanged model priors and seed, four chains and 2,000 warmup/retained draws per chain. The original failed run is retained for inspection. Computational checks do not certify forecast calibration.
2014 expanded runs · 2014 longer terminal run · 2017 expanded runs · 2014 baselines · Modern-cycle baselines
Reassessing shared polling error
The eligible final 2014 poll now ends six days before election day. Two pollsters contribute to the 2017 audit. The original cycle compositions are validated before combining minor parties solely for comparable error research.
| Held-out election | Training elections | Uncorrected | Training-mean correction |
|---|---|---|---|
| 2023 | 2020, 2017, 2014 | 1.56 | 1.48 |
| 2020 | 2023, 2017, 2014 | 1.45 | 2.23 |
| 2017 | 2023, 2020, 2014 | 1.92 | 1.96 |
| 2014 | 2023, 2020, 2017 | 0.85 | 0.78 |
The correction improves point error in 2 of four elections and worsens it in the others. It is not adopted. Four cycles still cannot identify an unrestricted six-dimensional error covariance or establish probability calibration. No shared-miss uncertainty scale has been selected.
Limits and reproducibility
Poll minus official result, in percentage points. Equal latest-poll-per-pollster ensemble; original rounded totals are normalised. Errors combine survey bias, sampling, turnout, undecided handling and movement after fieldwork. Same-direction errors are observations, not identification of a shared causal bias. Sparse cycles with uneven pollster coverage cannot establish calibration; full covariance is unidentified without strong regularisation. No fitted miss distribution or forecast probabilities are published. Every held-out training list excludes that election. Retrospective training on other cycles is distinct from operational past-only validation. Changing minor-party labels are collapsed only after validating the complete original cycle composition. Leave-one-election-out mean-error corrections are descriptive point benchmarks, not a fitted Bayesian miss prior. Held-out outcomes enter scoring only. Retrospective folds may include later cycles in training; they are not operational past-only forecasts. Check each row for latest fieldwork age and distinct eligible pollsters; records with unknown denominators remain excluded. No uncertainty scale was selected or calibrated.
Reid’s 2014 denominator and methodology need further primary documentation. The original final September 2017 Colmar report is still missing. Predictive and individual-prior checks on these final-cutoff inputs appear below. Earlier-horizon checks, electorate assumptions and seat scoring remain incomplete. Do not count repeated horizons as independent elections.
Download mappings, input hashes and held-out evidenceRECOVERED INPUTS · MODEL CHECKS
Which assumptions move the estimate?
We refit each final-cutoff model with one prior scale halved or doubled at a time. All runs use four chains and 2,000 warmup and retained draws per chain. Election results are used only after fitting for descriptive scoring; no prior is selected from these results.
| Prior / factor | Current-support shift | Election-day shift | Mean 90% interval width | Sampling |
|---|---|---|---|---|
| Reference / 1× | 0.00 | 0.00 | 4.12 | Passed |
| Opinion movement / 0.5× | 0.45 | 0.45 | 3.83 | Passed |
| Opinion movement / 2× | 0.49 | 0.48 | 4.58 | Passed |
| Pollster effects / 0.5× | Withheld | Withheld | Withheld | Failed |
| Pollster effects / 2× | 0.00 | 0.01 | 4.12 | Passed |
| Extra polling error / 0.5× | 0.12 | 0.13 | 3.98 | Passed |
| Extra polling error / 2× | Withheld | Withheld | Withheld | Failed |
2 of seven runs failed sampling diagnostics. Those comparisons are withheld; the full report preserves divergences and other diagnostics.
Can the reference reproduce its input polls?
These intervals describe replicated polls fitted by the model. The observations include the documented continuity correction and differ from the raw published percentages. High in-sample coverage does not establish predictive calibration.
Inspect replicated poll shares and uncertainty
morgan-2014-09-17
Maximum mean discrepancy: 2.03 percentage points.
| Party ID | Likelihood observation | Replicated mean | 90% range |
|---|---|---|---|
| National | 46.19% | 45.69% | 38.99–52.15% |
| Labour | 23.88% | 25.92% | 20.76–31.82% |
| Green | 13.47% | 12.46% | 9.21–16.09% |
| ACT | 0.58% | 0.76% | 0.22–1.77% |
| NZ First | 8.02% | 6.84% | 4.69–9.44% |
| Te Pāti Māori | 1.57% | 1.57% | 0.73–2.82% |
| Other / unreported | 1.08% | 1.25% | 0.50–2.49% |
| United Future | 0.58% | 0.76% | 0.22–1.80% |
| Conservative | 3.55% | 2.43% | 1.39–3.79% |
| Internet MANA | 1.08% | 2.32% | 1.01–4.34% |
morgan-2014-08-20
Maximum mean discrepancy: 2.01 percentage points.
| Party ID | Likelihood observation | Replicated mean | 90% range |
|---|---|---|---|
| National | 47.63% | 45.61% | 39.11–52.04% |
| Labour | 27.33% | 26.52% | 21.22–32.20% |
| Green | 11.49% | 12.18% | 8.89–15.97% |
| ACT | 0.59% | 0.77% | 0.21–1.84% |
| NZ First | 6.53% | 6.60% | 4.45–9.26% |
| Te Pāti Māori | 1.09% | 1.61% | 0.66–3.22% |
| Other / unreported | 1.09% | 1.24% | 0.48–2.47% |
| United Future | 0.59% | 0.78% | 0.21–1.88% |
| Conservative | 1.09% | 2.44% | 0.98–4.67% |
| Internet MANA | 2.57% | 2.25% | 1.20–3.71% |
morgan-2014-07-31
Maximum mean discrepancy: 2.80 percentage points.
| Party ID | Likelihood observation | Replicated mean | 90% range |
|---|---|---|---|
| National | 45.65% | 45.31% | 38.73–51.98% |
| Labour | 29.81% | 27.01% | 21.47–32.81% |
| Green | 11.98% | 12.12% | 8.78–16.01% |
| ACT | 0.59% | 0.77% | 0.20–1.84% |
| NZ First | 5.05% | 6.48% | 4.15–9.40% |
| Te Pāti Māori | 1.58% | 1.58% | 0.71–2.88% |
| Other / unreported | 1.09% | 1.25% | 0.47–2.51% |
| United Future | 0.59% | 0.78% | 0.20–1.84% |
| Conservative | 1.09% | 2.43% | 0.98–4.78% |
| Internet MANA | 2.57% | 2.28% | 1.20–3.84% |
These checks cover one cutoff in each of two elections. Earlier horizons, prior interactions, survey design effects and the initial-support prior remain unreviewed. Industry-wide miss remains prior-driven. Forecast publication stays blocked.
Reproducibility and remaining limits
Run: 2026-10-07T09:13:07.017597+00:00. Input version: bef9ebc10c8c9994ec32ae5f479cd8104beebeb4a51c2c203abb247ca03a88e1.
- In-sample predictive checks reuse fitted polls; they do not establish held-out polling or election-day calibration.
- One terminal cutoff per election; dependent parties and polls are not independent validation events.
- Cycles retain their original party categories. Do not pool their MAE or interval coverage.
- Individual changes do not test interactions. Initial prior, design effect, continuity correction, electorate assumptions and earlier horizons remain unreviewed.
- Shared-miss scale remains prior-driven. No alternative prior is selected from outcome scores.
The downloadable report includes full diagnostics, prior configurations, source input hashes, per-party predictive intervals and shared-miss sensitivity.
COMPUTATIONAL REVIEW · HELD-OUT POLLS
Predicting polls the model has not seen
Rechecking the three divergent runs
Each failed configuration gets one separately recorded retry with unchanged priors and seed, longer warmup and a smaller sampling step size. The original failures remain in the earlier report. Passing computational checks does not establish predictive accuracy.
| Election / configuration | Original divergences | Retry divergences | Retry R-hat | Sampling |
|---|---|---|---|---|
| 2014 / Tighter pollster-effect prior | 1 | 0 | 1.0033 | Passed |
| 2014 / Wider extra-error prior | 1 | 0 | 1.0032 | Passed |
| 2017 / Wider opinion-movement prior | 3 | 0 | 1.0032 | Passed |
Prediction before fieldwork begins
These four predictions use method v1. The newer v2 review above aligns prediction with the fitted likelihood directly. These archived scores are preserved and not pooled with the revised method.
For each target poll, training stops the day before its fieldwork starts. Target party results and election outcomes are absent from fitting and predictive draws. The published dates, pollster and gross sample size are used to evaluate the survey design. This is a poll prediction, with sampling and extra polling error; it is separate from an election-day forecast.
| Held-out poll | Training stops | Training polls | Model MAE | Latest-poll MAE | Mean CRPS | Party coverage in 90% intervals |
|---|---|---|---|---|---|---|
| morgan-2014-08-20 | 2014-08-03 | 1 | 1.10 | 0.70 | 0.67 | 10 / 10 |
| morgan-2014-09-17 | 2014-08-31 | 2 | 1.37 | 1.30 | 0.96 | 8 / 10 |
| colmar-2017-09-14 | 2017-09-08 | 1 | 3.01 | 2.85 | 2.05 | 6 / 8 |
| morgan-2017-09-16 | 2017-08-27 | 1 | 1.95 | 2.00 | 1.48 | 7 / 8 |
Among passing runs, the model has lower point error than the latest-poll baseline in 1 of 4 tests. This limited evidence does not justify adopting the Bayesian model over the public baseline.
Inspect held-out predictions and source cutoffs
morgan-2014-08-20
Training: morgan-2014-07-31. Target fieldwork: 2014-08-04 to 2014-08-17. Sampling: passed. House effects retain the training model’s centring.
| Party | Normalised published result | Likelihood observation | Predicted mean | 90% prediction interval |
|---|---|---|---|---|
| National | 48.00% | 47.65% | 43.75% | 32.46–53.83% |
| Labour | 27.50% | 27.34% | 29.03% | 20.16–38.56% |
| Green | 11.50% | 11.49% | 12.00% | 7.17–17.75% |
| ACT | 0.50% | 0.59% | 1.07% | 0.23–2.50% |
| NZ First | 6.50% | 6.53% | 5.27% | 2.70–8.47% |
| Te Pāti Māori | 1.00% | 1.08% | 1.95% | 0.69–3.83% |
| Other / unreported | 1.00% | 1.08% | 1.49% | 0.46–3.15% |
| United Future | 0.50% | 0.59% | 1.06% | 0.24–2.46% |
| Conservative | 1.00% | 1.08% | 1.48% | 0.45–3.06% |
| Internet MANA | 2.50% | 2.57% | 2.91% | 1.26–5.22% |
morgan-2014-09-17
Training: morgan-2014-08-20, morgan-2014-07-31. Target fieldwork: 2014-09-01 to 2014-09-14. Sampling: passed. House effects retain the training model’s centring.
| Party | Normalised published result | Likelihood observation | Predicted mean | 90% prediction interval |
|---|---|---|---|---|
| National | 46.50% | 46.21% | 45.69% | 38.94–52.39% |
| Labour | 24.00% | 23.89% | 27.93% | 22.35–33.99% |
| Green | 13.50% | 13.47% | 11.68% | 8.30–15.42% |
| ACT | 0.50% | 0.58% | 0.88% | 0.25–1.86% |
| NZ First | 8.00% | 8.02% | 5.93% | 3.79–8.43% |
| Te Pāti Māori | 1.50% | 1.57% | 1.58% | 0.68–2.88% |
| Other / unreported | 1.00% | 1.07% | 1.33% | 0.52–2.49% |
| United Future | 0.50% | 0.58% | 0.87% | 0.26–1.86% |
| Conservative | 3.50% | 3.55% | 1.32% | 0.52–2.48% |
| Internet MANA | 1.00% | 1.07% | 2.77% | 1.50–4.41% |
colmar-2017-09-14
Training: morgan-2017-08-18. Target fieldwork: 2017-09-09 to 2017-09-13. Sampling: passed. This unseen pollster receives a population-prior draw, not an estimated house-effect correction.
| Party | Normalised published result | Likelihood observation | Predicted mean | 90% prediction interval |
|---|---|---|---|---|
| National | 39.72% | 39.56% | 41.00% | 28.68–53.17% |
| Labour | 43.69% | 43.51% | 31.66% | 20.99–43.03% |
| Green | 6.95% | 6.98% | 9.24% | 5.07–14.62% |
| ACT | 0.60% | 0.67% | 1.05% | 0.25–2.45% |
| NZ First | 5.96% | 6.00% | 11.60% | 6.61–17.88% |
| Te Pāti Māori | 0.89% | 0.96% | 1.97% | 0.71–3.89% |
| Opportunity | 1.59% | 1.65% | 2.44% | 0.96–4.60% |
| Other / unreported | 0.60% | 0.67% | 1.04% | 0.24–2.39% |
morgan-2017-09-16
Training: morgan-2017-08-18. Target fieldwork: 2017-08-28 to 2017-09-10. Sampling: passed. House effects retain the training model’s centring.
| Party | Normalised published result | Likelihood observation | Predicted mean | 90% prediction interval |
|---|---|---|---|---|
| National | 40.00% | 39.81% | 40.90% | 29.80–51.09% |
| Labour | 39.50% | 39.31% | 31.71% | 22.16–41.82% |
| Green | 9.00% | 9.02% | 9.23% | 5.40–14.08% |
| ACT | 0.50% | 0.58% | 1.07% | 0.25–2.47% |
| NZ First | 6.00% | 6.04% | 11.61% | 7.00–17.28% |
| Te Pāti Māori | 2.00% | 2.07% | 1.97% | 0.74–3.81% |
| Opportunity | 2.00% | 2.07% | 2.44% | 1.00–4.46% |
| Other / unreported | 1.00% | 1.08% | 1.07% | 0.26–2.44% |
Four tests across two elections are too few to assess calibration. The tests share training polls, and party intervals are dependent. Do not pool these counts or treat broad coverage as proof of an accurate forecast.
Assumptions and reproducibility
- Only four pre-fieldwork held-out polls across two cycles; predictions share training data and are dependent.
- Historical archive is sparse and selected. This is not broad election-day calibration or an operational trained model.
- Gross sample size is an explicit approximation; target undecided and decided-base information is not used in prediction.
- Original cycle compositions differ; no pooled party-error or calibration statistic is published.
- A zero-divergence retry supports computational interpretation only, not model accuracy. Original failures are preserved.
- Held-out poll likelihood omits election-wide miss; this is a poll prediction, not an election prediction.
Run: 2026-10-07T09:25:03.141001+00:00. Report version: 9fabe46c769a4a97a4f31b844d159c95f2d6c830f2108cd810aa85e9b4d33c17.
Gross sample size is a proxy that can understate sampling error: published undecided figures and party-support bases are excluded from prediction. Sampling covariance uses predicted support rather than the held-out percentages. Remaining work includes design-effect sensitivity, shared-bias calibration and a broader archive.
HISTORICAL VALIDATION · ARCHIVED SNAPSHOT
Testing the 2014 party composition
This earlier snapshot contains four Bayesian fits using the ten categories recorded in 2014, including United Future, Conservative and Internet MANA. The earlier three-election reports below remain dated research snapshots. No forecast has been certified.
This archived run used two complete Roy Morgan releases, with final fieldwork 34 days before election day. It predates the recovered September 2014 Roy Morgan release and September 2017 Colmar Brunton PDF. See the expanded research above for updated coverage.
| Days before election | Latest fieldwork age at cutoff | Bayesian | Recency baseline | Sampling |
|---|---|---|---|---|
| 180 | — | Withheld | — | No complete exact-date model-eligible polls released by cutoff. |
| 90 | — | Withheld | — | No complete exact-date model-eligible polls released by cutoff. |
| 60 | — | Withheld | — | No complete exact-date model-eligible polls released by cutoff. |
| 30 | 4 days | 1.24 | 1.18 | Passed |
| 14 | 20 days | 1.27 | 1.18 | Passed |
| 7 | 27 days | 1.26 | 1.13 | Passed |
| 1 | 33 days | 1.26 | 1.13 | Passed |
All four fits pass convergence, effective sample size and divergence thresholds. In-sample predictive summaries are available in the report, but predictive review, prior sensitivity and election-day calibration remain incomplete. Seven cutoffs do not create seven independent elections.
Download the 2014 model runs, diagnostics and input hashes · Download the 2014 baseline comparisons
Does a shared-error correction generalise?
To compare changing party line-ups, this audit preserves six major parties and combines the remaining parties into one minor-party category, after validating each original composition. That aggregate is used only to compare polling errors; it cannot be used for seat allocation or threshold probabilities.
| Held-out election | Training elections | Uncorrected | Training-mean correction |
|---|---|---|---|
| 2023 | 2020, 2017, 2014 | 1.56 | 1.78 |
| 2020 | 2023, 2017, 2014 | 1.45 | 2.21 |
| 2017 | 2023, 2020, 2014 | 1.76 | 1.64 |
| 2014 | 2023, 2020, 2017 | 1.18 | 1.49 |
The simple correction worsens accuracy in 3 of 4 held-out elections. It is not adopted. Four election residuals have centred rank 3 across 6 compositional dimensions: they cannot identify an unrestricted shared-error covariance. No uncertainty scale has been selected or calibrated.
Inspect sources, assumptions and remaining gaps
Poll minus official result, in percentage points. Equal latest-poll-per-pollster ensemble; original rounded totals are normalised. Errors combine survey bias, sampling, turnout, undecided handling and movement after fieldwork. Same-direction errors are observations, not identification of a shared causal bias. Sparse cycles with uneven pollster coverage cannot establish calibration; full covariance is unidentified without strong regularisation. No fitted miss distribution or forecast probabilities are published. Every held-out training list excludes that election. Retrospective training on other cycles is distinct from operational past-only validation. Changing minor-party labels are collapsed only after validating the complete original cycle composition. Leave-one-election-out mean-error corrections are descriptive point benchmarks, not a fitted Bayesian miss prior. Held-out outcomes enter scoring only. Retrospective folds may include later cycles in training; they are not operational past-only forecasts. The final 2014 poll in this archive ends 34 days before election day and represents one pollster. No uncertainty scale was selected or calibrated.
2023 · 2 verified pollsters
1 NEWS Verian Poll · published 2023-10-11 · fieldwork ended 4 days before election day
Roy Morgan · published 2023-10-13 · fieldwork ended 6 days before election day
Official result2020 · 2 verified pollsters
Roy Morgan · published 2020-10-15 · fieldwork ended 20 days before election day
1 NEWS Colmar Brunton Poll · published 2020-10-15 · fieldwork ended 3 days before election day
Official result2017 · 1 verified pollster
Roy Morgan NZ telephone survey · published 2017-09-16 · fieldwork ended 13 days before election day
Official result2014 · 1 verified pollster
Roy Morgan · published 2014-08-20 · fieldwork ended 34 days before election day
Official resultDownload mappings, original input hash and held-out training evidence
EXPLORATORY VALIDATION
Historical baseline errors
Selected Roy Morgan, Colmar Brunton and Verian polls across three elections, with uneven source coverage. These are point-estimate errors, not evidence that a Bayesian forecast is calibrated. MAE is the average absolute party-share error in percentage points across all eight categories.
MAE 2.93 pp · RMSE 3.65 pp
1 polls from 1 pollsters · newest fieldwork 2023-03-26MAE 3.66 pp · RMSE 4.33 pp
1 polls from 1 pollsters · newest fieldwork 2023-06-25MAE 2.51 pp · RMSE 2.98 pp
1 polls from 1 pollsters · newest fieldwork 2023-07-23MAE 2.79 pp · RMSE 4.31 pp
1 polls from 1 pollsters · newest fieldwork 2023-08-27MAE 1.35 pp · RMSE 1.64 pp
1 polls from 1 pollsters · newest fieldwork 2023-09-26MAE 1.10 pp · RMSE 1.23 pp
1 polls from 1 pollsters · newest fieldwork 2023-10-03MAE 2.24 pp · RMSE 3.20 pp
1 polls from 1 pollsters · newest fieldwork 2023-10-08Coverage, assumptions and missing validation
Incomplete archive with uneven cycle and pollster coverage; 2017 has one verified pollster and two complete releases. Latest means most recently published, with fieldwork dates shown; contemporary publication and later article update dates may differ. Publication dates are respected; latest-poll baseline can be stale. Average uses all eligible polls ending within 45 days; recency uses 14-day half-life times square-root sample size (missing samples use an explicit 500 benchmark assumption). No tuned parameters, Bayesian fits, probabilistic calibration, seat prediction or electorate validation. Repeated horizons are dependent; do not count them as independent elections. All results are exploratory.
Bundled archive input version (live corrections require a new offline run): d17af83cf269d4dd · Benchmark baseline-v0.3
2017, 180 days: 0 released polls; 0 with fieldwork within 45 days. No eligible released polls.
2017, 90 days: 0 released polls; 0 with fieldwork within 45 days. No eligible released polls.
2017, 60 days: 0 released polls; 0 with fieldwork within 45 days. No eligible released polls.
2017, 30 days: 1 released polls; 1 with fieldwork within 45 days.
2017, 14 days: 1 released polls; 1 with fieldwork within 45 days.
2017, 7 days: 2 released polls; 2 with fieldwork within 45 days.
2017, 1 days: 2 released polls; 2 with fieldwork within 45 days.
2020, 180 days: 0 released polls; 0 with fieldwork within 45 days. No eligible released polls.
2020, 90 days: 1 released polls; 0 with fieldwork within 45 days. No recent polls for average/recency baselines.
2020, 60 days: 2 released polls; 1 with fieldwork within 45 days.
2020, 30 days: 3 released polls; 1 with fieldwork within 45 days.
2020, 14 days: 3 released polls; 1 with fieldwork within 45 days.
2020, 7 days: 4 released polls; 1 with fieldwork within 45 days.
2020, 1 days: 6 released polls; 3 with fieldwork within 45 days.
2023, 180 days: 1 released polls; 1 with fieldwork within 45 days.
2023, 90 days: 3 released polls; 1 with fieldwork within 45 days.
2023, 60 days: 5 released polls; 2 with fieldwork within 45 days.
2023, 30 days: 6 released polls; 1 with fieldwork within 45 days.
2023, 14 days: 7 released polls; 2 with fieldwork within 45 days.
2023, 7 days: 8 released polls; 3 with fieldwork within 45 days.
2023, 1 days: 10 released polls; 4 with fieldwork within 45 days.
Download the reproducible benchmark reportMODEL CRITICISM
Does the model earn its complexity?
Matched comparisons across 16 historical cutoffs and three independent elections. Simple averages sometimes outperform the Bayesian candidate. These results support further research, not a production-model choice.
2017 · 4 available cutoffs
- Bayesian mean absolute error
- 1.70 pp
- Latest-poll error
- 1.69 pp
- Recent-average error
- 1.52 pp
- Recency-baseline error
- 1.50 pp
Within-cycle average across matched available cutoffs. Lower is better; these are dependent comparisons, not independent trials.
2020 · 5 available cutoffs
- Bayesian mean absolute error
- 1.07 pp
- Latest-poll error
- 1.54 pp
- Recent-average error
- 1.48 pp
- Recency-baseline error
- 1.51 pp
Within-cycle average across matched available cutoffs. Lower is better; these are dependent comparisons, not independent trials.
2023 · 7 available cutoffs
- Bayesian mean absolute error
- 2.15 pp
- Latest-poll error
- 2.37 pp
- Recent-average error
- 2.33 pp
- Recency-baseline error
- 2.24 pp
Within-cycle average across matched available cutoffs. Lower is better; these are dependent comparisons, not independent trials.
Inspect every matched cutoff
| Election / days | Bayesian | Latest | Average | Recency |
|---|---|---|---|---|
| 2020 / 60 | 1.10 | 1.14 | 1.14 | 1.14 |
| 2020 / 30 | 0.97 | 1.64 | 1.64 | 1.64 |
| 2020 / 14 | 0.99 | 1.64 | 1.64 | 1.64 |
| 2020 / 7 | 1.06 | 1.75 | 1.75 | 1.75 |
| 2020 / 1 | 1.24 | 1.51 | 1.24 | 1.37 |
| 2023 / 180 | 2.85 | 2.93 | 2.93 | 2.93 |
| 2023 / 90 | 3.08 | 3.66 | 3.66 | 3.66 |
| 2023 / 60 | 2.52 | 2.51 | 2.17 | 2.20 |
| 2023 / 30 | 2.26 | 2.79 | 2.79 | 2.79 |
| 2023 / 14 | 1.59 | 1.35 | 1.91 | 1.56 |
| 2023 / 7 | 1.49 | 1.10 | 1.64 | 1.32 |
| 2023 / 1 | 1.29 | 2.24 | 1.20 | 1.22 |
| 2017 / 30 | 2.08 | 1.84 | 1.84 | 1.84 |
| 2017 / 14 | 2.07 | 1.84 | 1.84 | 1.84 |
| 2017 / 7 | 1.33 | 1.54 | 1.21 | 1.16 |
| 2017 / 1 | 1.32 | 1.54 | 1.21 | 1.16 |
Equal weight per election, then equal weight per available horizon within that election. Descriptive only; missing horizons differ by cycle. Baselines are no-change point predictions for election day; they have no probability distribution for a proper-score comparison.
How much do assumptions matter?
9 fresh fits test the final pre-election cutoff. The daily movement, pollster-effect and extra-error prior scales are halved or doubled together. This is a joint stress test: it cannot identify which individual prior causes a change. The election outcome is used only after fitting, for scoring.
Reference priors
Sampling diagnostics passed
- Largest current-support mean change
- 0.00 pp
- Largest election-share mean change
- 0.00 pp
- Mean 90% election-share interval width
- 5.05 pp
Half polling-prior scales
Sampling diagnostics passed
- Largest current-support mean change
- 0.24 pp
- Largest election-share mean change
- 0.28 pp
- Mean 90% election-share interval width
- 4.78 pp
Double polling-prior scales
Sampling diagnostics passed
- Largest current-support mean change
- 0.09 pp
- Largest election-share mean change
- 0.10 pp
- Mean 90% election-share interval width
- 5.19 pp
Separately, the shared election-wide miss prior was rescaled using the same draws at all 16 archived cutoffs. At this election’s final cutoff, mean 90% interval widths were 3.98 pp at 0.5×; 5.05 pp at 1×; 7.94 pp at 2×. This uncertainty is prior-driven and remains uncalibrated.
Can it reproduce the polling observations?
The archived log-ratio checks and fresh simulated-poll checks use the observations that fitted the model. High coverage can reflect wide intervals and reuse of the same data. It does not test election-day calibration.
100% of this poll’s eight modelled categories are inside their in-sample 90% predictive intervals. Largest absolute mean discrepancy: 1.47 pp.
NationalObservation 30.41%
Simulated mean 31.88% · 90% interval 28.03–36.05%
Simulated mean 31.88% · 90% interval 28.03–36.05%
LabourObservation 25.93%
Simulated mean 25.05% · 90% interval 21.52–28.69%
Simulated mean 25.05% · 90% interval 21.52–28.69%
GreenObservation 14.99%
Simulated mean 13.82% · 90% interval 11.22–16.65%
Simulated mean 13.82% · 90% interval 11.22–16.65%
ACTObservation 11.51%
Simulated mean 12.80% · 90% interval 10.18–15.72%
Simulated mean 12.80% · 90% interval 10.18–15.72%
NZ FirstObservation 7.53%
Simulated mean 6.48% · 90% interval 4.85–8.33%
Simulated mean 6.48% · 90% interval 4.85–8.33%
Te Pāti MāoriObservation 3.05%
Simulated mean 3.63% · 90% interval 2.30–5.30%
Simulated mean 3.63% · 90% interval 2.30–5.30%
OpportunityObservation 3.55%
Simulated mean 2.93% · 90% interval 1.93–4.19%
Simulated mean 2.93% · 90% interval 1.93–4.19%
Other / unreportedObservation 3.05%
Simulated mean 3.41% · 90% interval 2.21–4.91%
Simulated mean 3.41% · 90% interval 2.21–4.91%
Inverse log-ratio likelihood observations include the documented sample-size continuity correction. They are not a replacement for the preserved raw poll percentages. These are descriptive in-sample predictive intervals, not election-day calibration or held-out polling coverage.
Sampling checks and run provenance
R-hat 1.006 · bulk ESS 1479 · tail ESS 778 · divergences 0. Four chains; 1,000 warmup and 1,000 retained draws per chain. Run 2026-10-07T06:56:22.904508Z.
Input hash 15822d225d152c705b1b0f344388b33030594a20f2f91d80d5a8244f6fce868f. Outcomes never enter fitting.
Review scope, unresolved checks and reproducibility
Only three independent cycles; horizons, parties and repeated polls are dependent.
In-sample PPCs are not calibration tests. High coverage can reflect wide uncertainty and reuse of fitted data.
Baseline point predictions are interpreted as no-change election-day predictions. They have no predictive distribution, so probabilistic skill cannot be compared to them.
Equal-horizon summaries have uneven coverage; inspect matched per-cutoff rows. No significance test or winner is justified.
Industry-wide miss is prior-driven; sensitivity is not empirical calibration.
2014 uses another party composition; the contemporary secondary 2017 report is excluded.
No electorate reconstruction or seat forecast validation; no operational past-only trained model.
Zero-share correction, initial prior, design-effect assumptions and individual polling priors still require separate sensitivity checks before publication.
Review model-review-v1 · 2026-10-07T06:53:20.673988+00:00. Original historical fits retain their own timestamps and diagnostic flags. This descriptive review does not grant a publication pass.
Report hash 6a7fab0ff11684088d16ceb4289d317dc1ad50d3ebe719671834ad3bb929654d.
Retain the simpler Bayesian candidate for research. Do not add unsupported covariance complexity, select priors from these outcomes or publish probabilities. Broader source coverage and held-out calibration remain necessary.
Download the checks, sensitivity fits and matched baseline comparisonPSEUDO-REAL-TIME RESEARCH
Testing earlier forecast horizons
16 completed fits across 3 independent elections. 16 fits passed sampling checks. Two 2020 cutoffs are excluded because the election was postponed; three earlier 2017 cutoffs lack complete party data. This does not establish calibrated probabilities.
Data cutoff: 13 Oct 2023 · Election date known then: 14 Oct 2023.
10 eligible historical polls · 2 pollsters · Latest fieldwork 10 Oct 2023 (3 days old).
Primary election-date announcementVote-share error
- Bayesian mean absolute error
- 1.29 percentage points
- Latest-poll error
- 2.24 percentage points
- Recent-poll average error
- 1.20 percentage points
- Recency baseline error
- 1.22 percentage points
Sampling diagnostics
- Maximum R-hat
- 1.006
- Minimum effective samples
- 1479 bulk · 778 tail
- Divergences
- 0
Sampling checks passed. Limited predictive and prior stress checks are available in the model review; calibration and additional checks remain outstanding.
Audit this run
logistic-normal-rw-v0.3-isotropic-research · 4 chains · 1,000 warmup and 1,000 retained draws per chain. Fit completed 2026-10-07T06:05:19.530027Z. Historical outcomes enter scoring only, never inference.
Input hash: 15822d225d152c705b1b0f344388b33030594a20f2f91d80d5a8244f6fce868f
Used polls: morgan-2023-10-13, verian-2023-10-11, verian-2023-10-04, verian-2023-09-27, morgan-2023-09-05, morgan-2023-08-01, verian-2023-07-17, morgan-2023-07-04, morgan-2023-05-30, morgan-2023-03-31.
Official result used for scoringErrors average eight categories, including Other. Repeated horizons and party errors are dependent. Smaller errors on selected folds do not prove a better model. Common polling-miss uncertainty remains prior-driven; no seat-performance claim is made.
Download all horizons, proper scores, coverage gaps and run provenanceHISTORICAL MODEL RESEARCH
The first Bayesian backtest
These runs use polls available one day before the 2020 and 2023 elections. Fixed priors were not fitted to their election results. Sparse polling and two election outcomes cannot establish probability calibration.
2020 · Full covariance
Cutoff 16 Oct 2020 · 6 polls
- Mean party-vote error
- 1.14 percentage points
- Recent-poll average error
- 1.24 percentage points
- Maximum R-hat
- 1.050
- Minimum effective samples
- 136 bulk · 132 tail
- Divergences
- 0
Sampling checks did not pass. Scores are exploratory and cannot support model selection.
Run provenance and assumptions
Model logistic-normal-rw-v0.2 · 4 chains · 1000 warmup and 1000 retained draws per chain · seed 20261007.
Fixed regularising priors, selected terminal cutoffs only; not tuned to these election results. Scores from chains failing diagnostics are exploratory and unusable for model selection. Two elections cannot establish probability calibration. Polling coverage is sparse; no seat scores without as-known electorate assumptions. Industry-wide miss remains prior-driven. No historical results are used in inference, so this is a pseudo-real-time fixed-prior pilot, not a trained leave-one-election-out model.
Input hash: a5afafec74b9fe6f3f8225d96823d7dfc3b0d603bb8702cf9f4fbcdb50f27401
Published poll records: morgan-2020-10-15, colmar-brunton-2020-10-15, colmar-brunton-2020-10-08, morgan-2020-09-01, morgan-2020-08-08, morgan-2020-06-01.
2023 · Full covariance
Cutoff 13 Oct 2023 · 8 polls
- Mean party-vote error
- 1.63 percentage points
- Recent-poll average error
- 1.41 percentage points
- Maximum R-hat
- 1.570
- Minimum effective samples
- 7 bulk · 11 tail
- Divergences
- 0
Sampling checks did not pass. Scores are exploratory and cannot support model selection.
Run provenance and assumptions
Model logistic-normal-rw-v0.2 · 4 chains · 1000 warmup and 1000 retained draws per chain · seed 20261007.
Fixed regularising priors, selected terminal cutoffs only; not tuned to these election results. Scores from chains failing diagnostics are exploratory and unusable for model selection. Two elections cannot establish probability calibration. Polling coverage is sparse; no seat scores without as-known electorate assumptions. Industry-wide miss remains prior-driven. No historical results are used in inference, so this is a pseudo-real-time fixed-prior pilot, not a trained leave-one-election-out model.
Input hash: 947781a4a9344741665baf66eaa4e9acc4f22eee4818b41154878be2be1f23f2
Published poll records: morgan-2023-10-13, verian-2023-10-11, morgan-2023-09-05, morgan-2023-08-01, verian-2023-07-17, morgan-2023-07-04, morgan-2023-05-30, morgan-2023-03-31.
2020 · Simpler covariance
Cutoff 16 Oct 2020 · 6 polls
- Mean party-vote error
- 1.24 percentage points
- Recent-poll average error
- 1.24 percentage points
- Maximum R-hat
- 1.007
- Minimum effective samples
- 1302 bulk · 1405 tail
- Divergences
- 0
Sampling checks passed. Historical calibration and model review remain outstanding.
Run provenance and assumptions
Model logistic-normal-rw-v0.3-isotropic-research · 4 chains · 1000 warmup and 1000 retained draws per chain · seed 20261007.
Fixed regularising priors, selected terminal cutoffs only; not tuned to these election results. Scores from chains failing diagnostics are exploratory and unusable for model selection. Two elections cannot establish probability calibration. Polling coverage is sparse; no seat scores without as-known electorate assumptions. Industry-wide miss remains prior-driven. No historical results are used in inference, so this is a pseudo-real-time fixed-prior pilot, not a trained leave-one-election-out model.
Input hash: b4ba3e6a500a88feb0d84c7d1b72039c2280ecae5f6b29693e8fb06a3c5ab287
Published poll records: morgan-2020-10-15, colmar-brunton-2020-10-15, colmar-brunton-2020-10-08, morgan-2020-09-01, morgan-2020-08-08, morgan-2020-06-01.
2023 · Simpler covariance
Cutoff 13 Oct 2023 · 8 polls
- Mean party-vote error
- 1.26 percentage points
- Recent-poll average error
- 1.41 percentage points
- Maximum R-hat
- 1.009
- Minimum effective samples
- 832 bulk · 709 tail
- Divergences
- 0
Sampling checks passed. Historical calibration and model review remain outstanding.
Run provenance and assumptions
Model logistic-normal-rw-v0.3-isotropic-research · 4 chains · 1000 warmup and 1000 retained draws per chain · seed 20261007.
Fixed regularising priors, selected terminal cutoffs only; not tuned to these election results. Scores from chains failing diagnostics are exploratory and unusable for model selection. Two elections cannot establish probability calibration. Polling coverage is sparse; no seat scores without as-known electorate assumptions. Industry-wide miss remains prior-driven. No historical results are used in inference, so this is a pseudo-real-time fixed-prior pilot, not a trained leave-one-election-out model.
Input hash: 947781a4a9344741665baf66eaa4e9acc4f22eee4818b41154878be2be1f23f2
Published poll records: morgan-2023-10-13, verian-2023-10-11, morgan-2023-09-05, morgan-2023-08-01, verian-2023-07-17, morgan-2023-07-04, morgan-2023-05-30, morgan-2023-03-31.
A lower vote error on one date does not prove a model is better. Party errors and repeated horizons are dependent. Election-wide miss uncertainty remains driven by an uncalibrated prior; no seat-performance claim is made.
Download versioned research results, scoring and assumptionsUNCERTAINTY RESEARCH
When pollsters miss together
These are late-cycle poll results minus the official party vote. Positive means the poll overstated the result; negative means it understated it. Errors can include opinion change after fieldwork, turnout and survey bias.
1 NEWS Verian Poll
Fieldwork ended 4 days before election day. Released 11 Oct 2023.
National-1.59 pp
Labour+0.69 pp
Green+2.20 pp
ACT+0.23 pp
NZ First-0.17 pp
Te Pāti Māori-1.40 pp
Opportunity+0.05 pp
Roy Morgan
Fieldwork ended 6 days before election day. Released 13 Oct 2023.
National-7.58 pp
Labour-0.92 pp
Green+3.39 pp
ACT+2.86 pp
NZ First+1.41 pp
Te Pāti Māori-0.08 pp
Opportunity+1.28 pp
Parties with errors in the same direction across the verified pollsters: National, Green, ACT, Te Pāti Māori, Opportunity. This is an observation, not a probability or a diagnosis of its cause.
Verify the election resultWhy calibration remains blocked
3 independent election cycles, 7 compositional dimensions and residual rank 2. Repeated polls and forecast horizons do not create additional independent elections.
Poll minus official result, in percentage points. Equal latest-poll-per-pollster ensemble; original rounded totals are normalised. Errors combine survey bias, sampling, turnout, undecided handling and movement after fieldwork. Same-direction errors are observations, not identification of a shared causal bias. Sparse cycles with uneven pollster coverage cannot establish calibration; full covariance is unidentified without strong regularisation. No fitted miss distribution or forecast probabilities are published. Every held-out training list excludes that election. Retrospective training on other cycles is distinct from operational past-only validation.
Download the input-hashed audit and held-out training evidenceTHE OPINIONSCOPE BRIEF
Follow the next poll and forecast update.
New polls, meaningful movements and the evidence behind them. This is a launch list. Newsletter delivery has not started; a subscription will require email confirmation.
OpinionScope stores your request for up to one year. Contact the owner for access, corrections or removal.