Sealed on October 27, 2026, before the November 3 election. These are conditional forecasts of economic measures. They are not predictions of the election, of any vote or of any bill. This is research, not investment advice.
Elections can change economic policy, and that makes forecasting around them awkward. A forecast made before the vote has to assume an outcome; a forecast made after the vote already knows it. On October 27, 2026, ForeGlass, Comuvia's public forecast ledger, sealed a third kind: forecasts written in advance for each of several conditions, under a rule fixed before the vote that only the forecasts whose condition comes true will be graded.
What was sealed
The information cutoff was 14:15 UTC on October 27, 2026. Nothing published after it reached any forecaster or any brief. The forecasts cover three measures, each for two periods, the first half of 2027 and the calendar year 2027.
| Measure | What is forecast | Official series |
|---|---|---|
| Real disposable personal income | Growth of the period's average over the same period a year earlier | DSPIC96 |
| Unemployment-insurance benefits paid | Growth of the period's average over the same period a year earlier | W825RC1 |
| Federal budget balance | Sum over the period, in billions of dollars | MTSDS133FMS |
Real disposable personal income is income after taxes, adjusted for inflation. The federal budget balance is receipts minus spending, so a negative number is a deficit.
Every forecast is written under one of four conditions. None of the four is predicted, and no probability is attached to any of them.
- The House majority stays with the same party. The party with more seated members than all other parties combined when the 120th Congress convenes, scheduled for January 3, 2027, is the party that held the majority at the cutoff. At the cutoff that was the Republican Party: the House Press Gallery listed 218 Republicans, 214 Democrats, one independent and two vacancies.
- The House majority changes party. The first two conditions cannot both hold. If official records cannot settle the majority by February 5, 2027, neither is graded.
- S.2312 becomes law. The Unemployment Insurance Modernization and Recession Readiness Act was introduced on July 16, 2025 by Sen. Ron Wyden, a Democrat of Oregon. Among its provisions: a floor of 26 weeks of regular benefits for eligible claimants; a weekly benefit of at least 75 percent of the highest quarter's earnings divided by 13, up to a state maximum of at least two-thirds of the state's average weekly wage; and full federal funding of extended benefits, with stated exclusions.
- H.R.1119 becomes law. The Unemployment Integrity Act of 2025 was introduced on February 7, 2025 by Rep. Chuck Edwards, a Republican of North Carolina. Among its provisions: regular benefits conditioned on answering requests, attending interviews and reemployment services, and meeting other reasonable requests; voluntary reporting of noncompliance; a study of expanded audits within two years of enactment, with further audit changes depending on its findings; and transfers from a federal extended-benefits account conditioned on states certifying the new rules.
Each bill counts only if it becomes law as a bill of the 119th Congress, with the provisions named in the sealed condition substantially as introduced. That can happen no later than January 16, 2027. In the official bill-status records retained at the cutoff, the latest action on each bill was its referral to committee: the Senate Committee on Finance and the House Committee on Ways and Means. Each bill had six cosponsors.
What these forecasts cannot see
These limits come before any number, because they decide what any number can mean.
- No tested link from the conditions to the measures. No mechanism accepted after testing connects House control, or either bill's rules, to income, benefits paid or the deficit.
- No state detail. Unemployment insurance is run by the states; these forecasts are national.
- No reactions. No response of prices or of hiring to either bill is modeled.
- No later knowledge. Any update in January 2027 would know the election result, so it would be a new set of forecasts, graded on its own.
Who answered
The same forecasters answered every question.
- A declared naive baseline. It gives the no-change answer, assuming each measure stays where it last was, and takes an 80 percent range from its own past errors. It gives the same answer under all four conditions, because it carries no information about policy.
- The ForeGlass Engine abstained on all 24 questions. It forecasts only through mechanisms accepted after testing, and it has none linking House control or these bills to these three measures. The abstention is recorded as a result: it marks where the model's knowledge stops.
- Seven large language models, each asked twice. Once cold, without a brief, and once with an identical dated public brief pasted into the question. Both prompts also showed each question's no-change baseline value, and some questions' resolution rules gave a rough order of magnitude for the answer. None could search the web. Each gave a central value and a range it declared to hold the true value with 80 percent probability.
Three standard combinations of the models' answers form the ensemble forecasts: the median, which is the middle answer; the trimmed mean, the average after dropping the highest and lowest answers; and an inverse-variance weighted mean, an average that gives more weight to narrower ranges. The sealed record holds 496 forecasts across 24 questions: 24 from the baseline, 328 from the language models and 144 ensembles.
The 144 ensemble forecasts are public in the ForeGlass Verification Explorer. Type 2026-10-27 into its search box to count them; the table shows the first 120. Type house-retained, house-changed, s2312 or hr1119 to list one condition's 36. Where the ensembles differ between conditions, the difference comes from each model's single answer covering all four conditions, not from a measured effect of either bill or of House control. The individual model answers and the baseline's ranges stay in the sealed record, which the fingerprints below bind; the public ledger carries the baseline's central value for each question.
How the forecasts will be graded
- Conditions first, from official records. Each condition is assessed by its own route before any outcome is read. The House majority comes from the Clerk of the House when the 120th Congress convenes, with the party of the elected Speaker as the fallback. Each bill comes from the Government Publishing Office's bill-status record. Every assessment must be recorded by February 5, 2027, before any of the data being forecast exist, and none is ever decided by which forecast fits best.
- Forecasts under a condition that fails are void. Forecasts under a condition that official records cannot settle by February 5, 2027 are held instead. Both kinds are kept, never graded and never deleted.
- Forecasts under a condition that holds are graded on fixed data. Questions about the first half of 2027 are graded on the data as published by August 31, 2027; calendar-year questions on the data as published by February 29, 2028.
- The range and the central value are both graded. Each is checked against the published value: the range at the 80 percent coverage its forecaster declared, and the central value by how far it misses. Both are compared with the naive baseline on the same question. Misses are published exactly like hits. How ForeGlass grades a forecast explains the ledger's rules in general.
Limitations stated with the record
- Four conditions in one prompt. Each model answered all four conditions of a measure in the same prompt. Differences between its branches are a contrast within one answer, not four independent forecasts.
- A very wide baseline range. The baseline's range for unemployment-insurance benefits is very wide, because its past errors are dominated by 2009 and 2020. Landing inside it says little.
- Eight missing answers, named. One model's cold answers are missing for eight questions, real disposable income and unemployment-insurance benefits for calendar 2027 under all four conditions, because the provider was overloaded and the request failed. The record names the model, and the cold ensembles for those questions combine six models instead of seven.
- A case study, not a track record. This record sits outside every forecasting season, the pre-announced rounds used to measure skill. No claim of skill follows from it, and a branch that scores well says nothing about the branches that did not happen.
How to check it yourself
The record is anchored with OpenTimestamps, which ties a file's fingerprint to a Bitcoin block. An anchor proves that these exact bytes existed before that block was mined. It proves nothing about whether the forecasts are right, or about when anyone else could see them.
| What | SHA-256 fingerprint | Confirmed in Bitcoin block |
|---|---|---|
| The sealed design: conditions, questions, target data and the roster of forecasters | a94f908bb4a81045e829d887036457288a2eb72ef58ea27b708d0dcf5f9639e4 | 973414, mined October 27, 2026 at 15:14 UTC |
| The forecast registry with all 496 forecasts | 57a7529e944578152b10bacdb5677b88086dc28861bec12f1256f21d325d68a3 | 973423, mined 16:24 UTC |
| The public ledger with the 144 ensembles | 0cd9427902230371a62419dd47b5c4bed1dcd994fc7cbb096d188a6b923443cf | 973481, mined October 28, 2026 at 01:55 UTC |
The design's anchor comes before the registry's. The registry's fingerprint is anchored through a short text file that states it; that file's own fingerprint is 7b7be16b799ffbea80dd6371b0e1eb7c12744b7afd6db91e3bb3391636334224. The public ledger and its proof are at foreglass.ai/verification-explorer/ledger.json and ledger.json.ots. Drop both files on opentimestamps.org or run the open-source OpenTimestamps client to check them. The conditions themselves are bound to the sealed condition packet, fingerprint 6159356adcf5fd0f8b05b290a9c24983c77e08b3b5c5167f7f64de5b7ed9de50.
What this is not
It is not an election forecast. No probability is given for House control or for either bill. It makes no claim about what either bill or a change in House control would do: the ForeGlass Engine abstained, the baseline carries no policy information, and where the language models' forecasts differ between conditions, that difference comes from one prompt, not a measured effect. Both bills are described from their official records, with the same space and the same kind of detail.
Correction, 30 October 2026: an earlier version said each language model was asked "once cold" without saying what that prompt held. Both prompts also showed each question's no-change baseline value, and some resolution rules gave a rough order of magnitude for the answer. The description of the forecasters above now says so. No forecast changed.