Context, across elections.

These are conditional mathematical combinations, not government-formation predictions. No probabilities are assigned to the electorate scenarios.

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.

Proper scoring reference · Electoral Commission rules

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).

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.

Cutoffs use final official election dates retrospectively. They are a coverage audit, not operational forecasts using dates announced at the time.

• 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

• 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.

• 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.

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.

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.

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.

• 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.

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.

Download mappings, original input hash and held-out training evidence

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.

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.

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.

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.

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.

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.