The forecast record, checked November 4
Read the errors alongside the original rules. Election counts can change; provisional calculations are useful now, while final grades wait for the outcomes promised at publication.
The headline calculation
Custom normalized absolute-error score: 3.74. As of November 4, 2026. Election figures are provisional counts, not certified outcomes.
For a percentage or vote-share forecast, use |forecast − actual| in percentage points. For a dollar forecast, use 100 × |forecast − actual| / actual. Average eligible forecast errors within each issue, then give each included issue one equal weight. Lower scores mean smaller errors.
This is a custom composite mixing percentage-point errors and relative dollar errors. It is neither standard MAPE nor a pure percentage-point MAE, so the total has no percent sign. This methodology was selected after seeing outcomes on November 4, 2026. It is descriptive, not a preregistered validation test; future eligible forecasts will use the same formula and weighting.
| Issue | Included forecasts | Issue score | Scope and grading distinction |
|---|---|---|---|
| 01: School spending | 2 | 0.7837 | Both published spending runs, A and B, averaged within one issue. Registered Method B win remains unchanged. Shopping-start timing and deal-event planning do not have comparable outcomes. |
| 04: September non-Senate elections | 131 | 10.3483 | 131 provisional named-pair, ballot and retention errors from DDHQ, with a later WCAX Vermont update. Excludes 29 overlapping Senate forecasts, 3 candidate mismatches and 62 unresolved survey/map items. Vermont Proposal 4 uses a later WCAX count at 71.4% reported; Wyoming's registered named pair omits a candidate who currently has more votes than one of the pair. Certified winner-call grades remain pending. |
| 06: Household spending growth | 1 | 0.1953 | Published headline forecast; August 2026, Data!B42. Original binary loss: within tolerance, but persistence was closer. The later display label is Partial. Thirteen secondary measures are shown separately. |
| 07: Holiday spending | 1 | 3.9299 | Published headline forecast. Original binary loss: within tolerance, but persistence was closer. The later display label is Partial. Two expectations are shown separately. |
| 11: October Senate forecast | 35 | 3.4286 | Provisional DDHQ counts; named-candidate two-way shares. Includes Alaska first-choice votes as an explicitly provisional proxy; the registered grade requires its final ranked-choice round and all 35 certified results. |
Issues 6 and 7 contribute only their published headline measures. Their secondary results remain visible on the issue pages and do not change headline weighting. Issue 1 retains both published spending forecasts within its single weight. Issue 11 replaces overlapping September Senate forecasts; Issue 3 is reported separately and excluded from this composite because its races are reforecast later.
Spending inputs and sources
- Issue 1: $856.83 and $870.37 versus $863.86. Retrieved 2026-11-04T17:43:43Z. Both published spending runs, A and B, averaged within one issue. Registered Method B win remains unchanged. Shopping-start timing and deal-event planning do not have comparable outcomes.
- Issue 6: 4.9% versus 4.704714298%. Retrieved 2026-11-04T17:33:12Z. Published headline forecast; August 2026, Data!B42. Original binary loss: within tolerance, but persistence was closer. The later display label is Partial. Thirteen secondary measures are shown separately.
- Issue 7: $1589.00 versus $1654.00. Retrieved 2026-11-04T17:34:34Z. Published headline forecast. Original binary loss: within tolerance, but persistence was closer. The later display label is Partial. Two expectations are shown separately.
Exclusions are about comparability and duplicate targets, not whether a forecast was right. Issue 2 is a categorical council decision, not a 94.9% measured public-support outcome. Issue 5 is retained as a miss in index points; no conversion for index scales is defined by this composite. Issues 8, 9 and 10 await their registered releases. Survey timing, population and instrument mismatches stay outside the calculation until resolved.
The previous October 19 spending MAPE was 2.9548%. This November 4 composite uses a different formula and adds provisional election issues, so the two totals do not measure the same thing.
Registered win/loss rules remain separate. Issues 6 and 7 originally lost their strict baseline tests, even though both met their tolerances; their later “Partial” labels do not turn those losses into wins. The election issues retain pending registered grades until the specified final results are available.
Read the complete November 4 audit, pending releases and comparison limitations.
Every issue
Issue 01 · Method B spending win retained
NRF July 14: $863.86. Method A $856.83; Method B $870.37. The composite averages both dollar-relative errors. Shopping-start timing is an instrument mismatch; the later matched amendment is descriptive, not a replacement primary. The deal-event planning forecast lacks a comparable outcome.
Issue 02 · Categorical win retained
The recorded county-council action is a direction match. It does not establish accuracy for the forecast public-support percentage or provide a numeric outcome for this composite.
Issue 03 · Registered grade pending; provisional results available
All five July Senate races are shown against the same DDHQ snapshot used for Issue 11. North Carolina and Maine remain labeled diagnostic cases. All five are excluded from the composite as later reforecasts.
Issue 04 · Row-level election and survey review
All 225 original questions remain frozen. Candidate-lineup mismatches, overlapping Senate targets, source verification holds and registered future survey waves are distinguished on the issue page. Provisional share errors do not constitute certified winner-call grades.
Issue 05 · Registered primary miss
September preliminary ICS: 55.4 versus 47.8, error 7.6 index points. ICC: 61.8 versus 50.9, error 10.9. ICE: 51.4 versus 45.8, error 5.6. The secondary ICS companion 52.1 has error 4.3 and cannot replace the primary. The September final release is not the registered target.
Issue 06 · Original loss; later Partial display
Headline 4.9% versus 4.704714298%, error 0.195285702 percentage points. All thirteen secondary measures have workbook-matched descriptive comparisons. Only the headline enters the composite.
Issue 07 · Original loss; later Partial display
$1,589 versus $1,654: $65 absolute error, or 3.9299% of the outcome. Higher-price expectation 83.3% versus 79%: 4.3 percentage points. The economy-weaken comparison remains conditional because the complement of same-or-improve is not verified as an exhaustive weaken category.
Issue 08 · Pending November 9 release
October SCE unemployment expectation, 47.33%. Await the full-precision registered survey outcome; this is not a forecast of the unemployment rate.
Issue 09 · Pending November 16 release
October SCE Credit Access: auto 12.10%, card 22.86%. Both predictions will be scored and averaged within one issue when comparable outcomes exist.
Issue 10 · Pending October survey release, expected November 10
NFIB business health: Excellent 14.2%, Good 43.6%, Fair 32.3%, Poor 10.0%. A release labeled October but reporting September data is not the target.
Issue 11 · Registered grade pending; provisional results available
All 35 Senate races use a saved DDHQ result vector with per-race source timestamps. Alaska is first-choice only; its final ranked-choice round and certification remain outstanding. Numerical errors and source-supported winner calls are reported separately.
Comparisons with other forecasts
On the same 35 provisional Senate counts, the average absolute error in named-pair vote share is 3.43 points for Lewsearch and 4.21 points for Synthetic Users. We normalized the same two named candidates in each organization’s forecast and in the vote count; Synthetic Users’ published shares are rounded to one decimal before this calculation. Its desk was updated November 2, while Lewsearch’s forecast was frozen October 31. Alaska uses first-choice votes here, not the final ranked-choice round; Alaska and Mississippi were uncalled. See the race snapshot and comparison.
A margin shortcut, 50 + signed margin / 2, does not always recover a named-pair share when other candidates or undecided responses are present. The Montana and Idaho pages illustrate the difference; the comparison above uses the displayed candidate shares rather than that shortcut.
We do not recommend a competitive headline from this comparison. Thirty-five races, different publication dates, correlated errors and changing counts limit the conclusion. The terms reserve intellectual property rights, and the acceptable-use policy restricts systematic extraction from the Services. These terms do not establish a clear license to republish the forecast dataset. A factual attribution and source link should not suggest endorsement; obtain appropriate permission before reproducing their dataset or branding.
A fresh search found Prolific’s November 3 national House generic-ballot poll, but no verified comparable state-by-state Senate forecast. Its weighted likely-voter Democratic margin is 9.3 points; its dynamic MRP estimate is 9.8. These are separate estimates. They can be evaluated separately against a sufficiently complete national House popular-vote margin, with the limits of a generic-ballot-to-election comparison disclosed, including uncontested districts. They cannot be entered into a Senate race MAE.
Provisional results and next updates
These dated calculations show the outcomes available now, with their denominators, sources and unresolved cases. Final election grades await the outcomes specified in each issue’s registered rules. Later updates will retain misses, the registered primary forecasts and the target survey waves.
Release calendars: New York Fed; NFIB Small Business Economic Trends. Source snapshots and calculation inputs are retained with the grading data.