Prediction markets priced 10 midterm questions the night before. All 10 favourites won. That proves less than you think.
On the eve of the US midterms (read at 00:41 UTC on 3 November 2026) we froze the prices of ten Polymarket markets on Congress. Yesterday they resolved. Every market's favourite won.
Here is the whole record, eve price beside the result, as MonitoringRoom now shows it on the page:
| market (Polymarket) | priced on the eve | resolved |
|---|---|---|
| Democrats control the House | 94.5 % | YES |
| Republicans control the House | 5.5 % | NO |
| Democrats control the Senate | 75.5 % | YES |
| Republicans control the Senate | 24.5 % | NO |
| Balance of power: D Senate, D House | 75.5 % | YES |
| Balance of power: R Senate, D House | 18.5 % | NO |
| Balance of power: R Senate, R House | 5.5 % | NO |
| Balance of power: D Senate, R House | 0.4 % | NO |
| Balance of power: Other | 0.1 % | NO |
| Republican trifecta with a Senate supermajority | 0.5 % | NO |
Outcomes were read from Polymarket's own market records on 5 November 2026 (00:46 UTC). The eve prices were frozen two days earlier and never rewritten.
"The markets were right" is the wrong lesson
Ten for ten sounds like a perfect night. It is less impressive than it looks, and the reason is the most useful thing to know about reading odds.
Most of these were not hard calls. Six of the ten were priced at 5.5 % or less, or 94.5 % or more. A market that says 95 % is supposed to be right 19 times in 20. Getting those right is the expected result, not a feat.
The Senate was the real question. At 75.5 %, the market was saying: Democrats take the Senate about three times in four. That also means the other outcome should happen about one time in four. If you saw ten different 75 % calls and all ten came true, you should start to suspect the prices were too low, not celebrate.
One night cannot show a market is well calibrated. Calibration means that across many questions, things priced at 75 % happen about 75 % of the time. You learn that from hundreds of resolved markets, not from one election.
A score that is honest about it
The standard way to grade probability forecasts is the Brier score: for each question, take (price − outcome)², with the outcome as 1 or 0, and average. 0 is perfect. Saying "50 %" to everything scores 0.25.
eve = [0.945, 0.055, 0.755, 0.245, 0.755, 0.185, 0.055, 0.0045, 0.0005, 0.0055] won = [1, 0, 1, 0, 1, 0, 0, 0, 0, 0] print(sum((p - o) ** 2 for p, o in zip(eve, won)) / len(eve)) # 0.0223... 0.022 is a very good score. About 80 % of the penalty comes from just three markets, the ones that hung on the Senate at 75.5 % / 24.5 %: each still costs 0.06 even though the favourite won, because the market left a one-in-four chance on the table. That is the point. A good forecaster is not someone who is never surprised. It is someone whose 75 % calls come true about 75 % of the time.
How to read the next one
- Read the price as a frequency. 62 % means "about 6 times in 10", not "probably yes".
- Look at the close calls. The 25–75 % markets are where the information is. The 95 % ones mostly confirm what everyone knew.
- Keep the eve price. Prices move as results come in, so the only fair test is the price before anyone knew. That is why we froze it.
The page keeps the eve price beside each result for good: monitoringroom.com/us-midterms-2026. There is also a plain guide to how prediction-market odds work.
Information only — not betting or political advice. MonitoringRoom reads public market prices; it takes no positions and sells no picks.