Election night
Commission polls, call the race, then see why a 2-point lead is often a coin flip.
Commission a poll. 8 of 8 left tonight.
- Chips
- …
- play chips, not money
- Score
- –
- after your first call
- Right
- 0
- calls on the winner
- Streak
- 0
- best 0
Latest poll
A hidden share of voters backs A. Poll up to 8 times, then say how sure you are that A wins. Chips reward being right about how sure you are, not just being right.
50/50: no chips either way.
Run at least one poll before you lock in.
How often were your 70% calls right?
Your report appears after 10 races. 10 to go.
Being calibrated means your 70% calls come true about 70% of the time. Right more often than that and you could have said a bigger number. Right less often and you were too sure.
The Brier score is the squared gap between what you said and what happened, averaged over races. Always saying 50% scores 0.25. A perfect seer scores 0. Saying exactly what you believe earns the most chips on average, so honesty pays.
Your report appears after 10 races. 10 to go.
A thousand elections
Give A a real lead and count how often the polls name the wrong leader.
A 51.0% · B 49.0%
- One poll
- ?
- named the wrong leader · formula 26.4%
- Average of 5
- ?
- named the wrong leader · formula 7.9%
- Average of 10
- ?
- named the wrong leader · formula 2.3%
A poll of 1,000 has a margin of ±6.2 points on the lead. Try a 2-point lead with 1,000 people, then switch to 2,500.
The real one
US midterms, Tuesday, November 3, 2026: what the markets and the polls say.
What the markets say
Who controls the House
Polymarket: 95% chance Democrats control the House, 5% chance Republicans do.
Kalshi: 94% chance Democrats control the House, 6% chance Republicans do.
Who controls the Senate
Polymarket: 71% chance Democrats control the Senate, 29% chance Republicans do.
Kalshi: 72% chance Democrats control the Senate, 28% chance Republicans do.
How to read it: a market trades a contract worth $1 if something happens. A price of 67 cents means traders, together, put the chance at about 67%. Each number here is the middle of the current price range, scaled so the two parties add to 100%.
Polymarket, Oct 31, 2026, 1:15 PM ET · Kalshi, Oct 31, 2026, 1:15 PM ET
Shown to teach probability, not betting advice.
What the polls say
Democrats +8.6
The average of 5 poll averages for the generic congressional ballot.
- Decision Desk HQD +8.9
- FiftyPlusOneD +8.8
- NBC NewsD +8
- RealClearPoliticsD +8.2
- Silver BulletinD +8.9
Each average as listed on Wikipedia, checked Oct 31, 2026, 1:15 PM ET. Behind them: 10 recent public polls in VoteHub's free data, the newest from 2026-10-04.
How to read it: the generic ballot asks which party you would vote for in your House district. It measures the national vote share, not seats. Other averages to compare: the list above. When averages built different ways agree, trust them more.
The Senate has no national poll to read. Only a third of the seats are up, and each one is its own state race, so read state polls one at a time.
Turn a poll lead into a chance
If the lead is roughly normal around the true lead, the chance the leader is really ahead is Φ(lead ÷ SE), where SE is the margin of error on the lead divided by 1.96.
±6.2 is one poll of 1,000 people: ±3.1 on each share, doubled for the lead. An average of many polls has a smaller one.
Polls, if only luck mattered: 99.7%.
- Polls, if only luck mattered
- 99.7%
- chance the leader truly leads the vote
- Markets, House control
- 94 to 95%
- chance the same party wins the House · Polymarket 95% · Kalshi 94%
To get the markets’ 94 to 95% from a 8.6-point lead, the margin on the lead would have to be about ±10.8 points.
Why they disagree. A poll measures today’s vote share, and its margin only counts the luck of who got asked. A market prices the final result, so it also weighs turnout, how district lines turn votes into seats, and the chance that every poll misses the same way. Winning the national vote is not the same as winning the House. That is why a 2-point poll lead is not a 2% edge, and why a big lead is still not a sure thing.
What’s going on
The real idea behind the game, in plain words.
A poll of n people estimates a share p with a 95% margin of error of about 1.96 × √[p(1 − p)/n]. Near a 50/50 race that is close to 1/√n: about ±4.9 points for 400 people, ±3.1 for 1,000 and ±2.0 for 2,500. To halve the margin you need four times the people.
That headline margin is on each candidate’s share. The lead, A minus B, has about twice the margin, because a voter who switches adds to one side and takes from the other. A poll of 1,000 is good to about ±6 points on the lead, so a 2-point lead sits well inside the noise. With a true 2-point lead, one poll of 1,000 shows the wrong leader about 26% of the time.
Averaging polls works like one bigger poll. Five polls of 1,000 act like a single poll of 5,000, so the margin shrinks by √5 and the wrong leader rate falls to about 8%. Ten polls bring it near 2%.
Saying how sure you are is scored with the Brier score, (outcome − p)², where outcome is 1 if A wins and 0 if not. It is a proper scoring rule: your best average score comes from saying exactly what you believe, so shading a 70% hunch up to 95% costs you over many races. Chips are the same score moved so that 50% pays nothing: 100 × (1 − Brier) − 75.
The formula only counts sampling error, the luck of who gets picked. Real polls also miss because of who answers, who actually turns out to vote and how questions are worded. Those errors tend to push every poll the same way, so averaging cannot cancel them. That is why real elections miss the polls by more than the margin of error says.