Polymarket 2026 midterm scorecard: 13,186 contracts frozen before the vote

At the freeze, 402 of 507 Senate, House and governor races had a favourite priced at 90¢ or more, 332 at 95¢ or more, and only 10 below 60¢. Each race is counted once by its highest-priced contract, read as the midpoint of the best bid and ask in Polymarket's public catalogue at 03:34 UTC on 17 October 2026.

That is the shape the scorecard has to be honest about. Most contracts are near-certain and will score well whatever happens; the test is in the 105 races priced under 90¢ and in how often the favourites in each price band actually won. Both tables are declared below, before any result exists.

Download chart figures:CSVJSON with sourcesSVG chart

The universe is every market of every event carrying Polymarket's `midterms` tag, open or already closed, plus 45 events found by a title screen of 22 related tags and the whole 22,858-event active catalogue and confirmed by hand: California's governor, lieutenant governor and propositions, state ballot measures, and Senate-and-governor combinations. The sweep finished at 03:34:08 UTC on 17 October 2026. Every discovered contract keeps a row, scored or not, with its ids, rule text and capture time.

Tier A is the headline set: race winners and chamber control. Tier B is everything else the general election settles. Tier C stays listed and unscored: 1,058 contracts already determined before the freeze, mostly primaries, and contracts that something other than an election outcome settles, such as a forecaster's odds, a campaign event, a leadership vote or a combination with an oil-price leg. 2,621 lettered placeholder outcomes keep their rows but are left out of the Tier A means.

Download table figures:CSVJSON with sources

The forecast that is scored is the order-book midpoint of every contract at 23:59:59 UTC on 2 November 2026, the night before the vote. A second checkpoint at the same hour on 27 October measures what the final week added. Captures run hourly from 25 October and every five minutes around both instants; the checkpoint reading is the latest capture that finished before the instant whose order-book requests answered for at least 99% of the manifest's tokens; a token the venue holds no order book for is an answered request and becomes a contract without a quote. A capture that finished after the instant is never used, whatever it shows, and no reading is carried forward.

The manifest's SHA-256 begins 5b607bde and is printed in full below. It covers the canonical JSON of the events and contract rows without the capture time, so anyone holding the file can recompute it. Prices are not part of the manifest. Contracts Polymarket lists after the freeze are recorded by each capture's re-sweep and scored in a separately versioned supplement, never merged into version 1.

5b607bded8bd47c8f41ec083f54d6319d7178027c7914c64ab77f069c4a79dca

Each contract gets a Brier score, the squared gap between its checkpoint price and what happened. The price is the midpoint of the best bid and best ask of the Yes order book; a contract with no bid or no ask has no price, stays in every table and is counted as missing. A spread wider than 10¢ keeps its price with a flag, and Tier A is reported with and without those rows.

Means are reported two ways: over contracts, and over races with every race weighted equally, because a race with twelve contracts is still one result. For every race the highest-priced contract is its favourite, and a table shows, per 10¢ band, how many favourites there were and how many won. Calibration bins of 10¢ report count, mean price and share resolved Yes. Two baselines use the same contracts and clocks: a constant 50¢, and the 27 October price. Intervals come from a bootstrap that resamples races, not contracts, 2,000 times from a fixed seed. No comparison with polls, forecasters or another venue is made, because none has a matched, rights-cleared benchmark. One election is one election: the text will say how little independent information it supplies.

Download table figures:CSVJSON with sources

Every contract gets two outcome records that are never merged. The official record is the result published by the state or territorial election authority, first as an unofficial count and later as certified totals, with its URL and retrieval time; an Associated Press call is logged as an interim status, not as the official record. The venue record is Polymarket's own payout, read from its public market record with its resolution date and any dispute. The official record decides the score wherever it can; a disagreement between the two is flagged and listed, and a question no election authority answers is scored on the venue record and marked as such.

Voided or split resolutions leave every mean and are counted. A contract without a result at a scoring run is pending: shown with its price, left out of that run's means, filled in a later dated run that never rewrites an earlier one. The first run is planned for 4 to 6 November 2026 and will be labelled by status; certifications, recounts and the Georgia-style runoffs come in later runs. A contract whose rules changed between the freeze and a checkpoint is scored on the frozen text and listed.

Download the frozen manifest as CSV (35 MB, one row per contract with its rule text) or as gzipped JSON (events and rows, with the hash rule inside), the Gamma fields seen at the freeze, the rules file, the methods note, the coverage report and the code for the sweep, the captures and the scoring run. The file list carries a SHA-256 for every published file.

After the election this page adds the scored tables, the per-contract rows with both outcome records, the capture receipts and a dated corrections list. The result is not known and is not on this page. Related Vultax work: the monthly Polymarket efficiency series and the study of how long Polymarket takes to settle, which is why pending contracts are expected after 3 November.