The polls promise Democrats a big Tuesday. The polls have failed them before.

The surveys say blue. The calendar says beware.

Americans vote on Tuesday, November 3, in midterm elections that will decide all 435 seats in the House of Representatives and 35 of the 100 Senate seats; and nearly every public poll says Democrats are heading for a good night. The RealClearPolling average of the national congressional ballot gives Democrats an 8.2-point lead; Silver Bulletin puts their chances of taking the House at 95.1 percent and the Senate at 77.7 percent, while Decision Desk HQ gives them 80 and 60 percent respectively.

The eye-catching details come from the states. A New York Times/Siena poll found appointed Republican Senator Darline Graham and Democratic challenger Annie Andrews tied at 48 percent in South Carolina; a state Donald Trump carried by 18 points in 2024. CNN's final battleground surveys put Democrat Abdul El-Sayed ahead of Republican Mike Rogers 49 to 42 in Michigan and Democrat James Talarico ahead of Republican Ken Paxton 49 to 43 in Texas, with no clear leader in Maine or Kansas.

Democrats lead in well over the 218 seats needed to flip the House, the polls suggest, and run tied or narrowly ahead in the Senate races that matter. The catch, one Democrats keep repeating to themselves, is that they have heard all of this before.

What the numbers actually say, state by state

The starting balance: Republicans hold 218 House seats to the Democrats' 214, with one independent and two vacancies; in the Senate, Republicans hold 53 seats against 45 Democrats, with two independents who align with the Democrats. Democrats need 218 for a House majority and a net gain of four Senate seats for outright control, since Vice President JD Vance would break a 50-50 tie for the Republicans.

On the generic ballot, which party voters say they will back for Congress, Democrats lead by 8.2 to 8.3 points in the RealClearPolling average, with 51.4 percent against 43.2 percent. Forecasters translate that into seats: Silver Bulletin's October 29 model projected an average of 240 Democratic House seats and a 53-47 Senate; Decision Desk HQ projected 238 and 52; the Cook Political Report foresaw Democratic gains of 12 to 22 House seats. Prediction markets are even more confident: Polymarket priced a Democratic House at 93 percent and a Democratic Senate at 68 percent.

Beneath the national numbers, the state polls describe an expanded map: the Times/Siena surveys showed dead heats in South Carolina and Kansas and a 54-to-40 Democratic lead for Peggy Flanagan in Minnesota, while CNN found Democrats numerically ahead in all four of its final battleground states.

The questions the models cannot answer

What the polls cannot tell us is who will actually show up. Late-deciding voters, still a meaningful slice of the electorate in several Senate races, could break uniformly in one direction, as they appear to have done in Wisconsin's August primary. Nor is it settled whether Republican voters are being systematically undercounted: Ipsos Public Affairs president Clifford Young told the BBC that conservatives, particularly on the far right, are "much more anti-establishment by nature" and refuse to take part in surveys at all, which could skew models toward Democrats if pollsters fail to adjust for it.

It is also unclear how much weight to give any single state poll. The South Carolina tie comes from one high-quality survey of 707 likely voters with a 4.3-point margin of error; Fox News's forecaster moved the race only to "toss-up" while the Cook Political Report still rates it Lean Republican. And nobody knows what early-vote figures truly mean; they reveal turnout, not choices.

In short: the direction of the polling averages is consistent, but the electorate that produces Tuesday's result may not be the electorate the models imagined.

Two years of Trump's presidency are on the ballot

If the polls are right, Democrats retake the House and with it the power to set the legislative agenda, launch investigations and block the White House for the final two years of Donald Trump's second term; a term in which, as this newsroom has reported, the war with Iran has itself become a midterm issue. A Democratic Senate would go further, deciding the fate of judicial nominations and confirmations.

If the polls are wrong in the other direction, the consequences are institutional rather than partisan: a third consecutive cycle of meaningful polling error would detonate what remains of the survey industry's credibility, call the forecasters' models into question, and leave prediction markets, which have priced a Democratic sweep as the near-certain outcome, nursing heavy losses. Either way, Tuesday will also render a verdict on the Democratic Party's own divisions, laid bare this weekend when Representative Haley Stevens, fresh off losing Michigan's Senate primary, crossed party lines to endorse Republican Mike Rogers.

The battle for Congress is about arithmetic; it is also about whether either party can govern what it wins.

How did the pollsters get the 2026 primaries so wrong?

Start with Wisconsin. In August, Milwaukee County Executive David Crowley defeated state Representative Francesca Hong in the Democratic primary for governor by about 3,700 votes, half a percentage point, after polls had shown Hong leading by as much as 20 points. The final Marquette Law School poll was partly conducted while Crowley was out of the race entirely, having briefly suspended his campaign before re-entering with Governor Tony Evers's endorsement. Late deciders broke sharply for Crowley in a way no public survey captured.

Then Michigan: polls suggested Abdul El-Sayed was headed for a comfortable, even double-digit win in the Democratic Senate primary; he won by less than a percentage point, according to the Wall Street Journal; and the bad blood never healed, culminating in one of the cycle's ugliest Democratic feuds and Stevens's Sunday endorsement of Rogers. Add the revelation, reported by the Washington Examiner, that the pollster Wedgewood admitted fabricating survey data across Senate and gubernatorial races; its invented numbers circulating on aggregators for weeks.

The industry's defense, offered by Ipsos's Young to the BBC, is structural: Americans have abandoned the landlines pollsters once relied on, the hardest voters to reach are the most anti-establishment, and humans, his word, "are not easy." The counter-evidence is 2022, when midterm polls performed better, partly because Trump was not on the ballot to scramble turnout models. The question for Tuesday is which precedent holds.

What a polling miss would detonate beyond Tuesday

Consider the two failure modes. If Democrats win roughly what the polls predict, a House majority in the 230s, a narrow Senate, the story is a conventional wave against an unpopular president presiding over rising food and bill costs, and Washington returns to divided government. If Republicans hold on, the detonation is epistemic: pollsters, forecasters and markets would have misread the electorate in three straight presidential-era cycles, and the "calcified electorate" thesis, fewer swing voters, deeper polarization, would need rewriting.

Either outcome reshapes 2028. A Democratic Congress constrains Trump's final two years and elevates a new generation of committee chairs; a Republican hold, against every model, would be read as proof that the party's coalition now systematically evades measurement. There is a quieter second-order effect too. The Stevens defection in Michigan, a sitting congresswoman endorsing the Republican two days before the vote, then being disowned by her state party, shows a Democratic coalition that polls describe as unified behind its nominees but that behaves, in places, like anything but.

What should you watch on Tuesday night?

Watch the House first: with Democrats needing only a handful of net gains, early calls in competitive districts will signal whether the night follows the models. Then the Senate map, roughly in order of poll-closing time: the dead heats in South Carolina and Kansas, the CNN battlegrounds of Michigan, Texas and Maine, and the longer-shot Democratic hopes in Alaska, Iowa and Ohio.

Treat early-vote dumps with caution, they describe who voted, not how, and remember that undecided voters breaking one way late is precisely the scenario the models handle worst. Note the recount tripwires: Wisconsin's rules offer a recount below one percent, the exact band in which Crowley beat Hong. And keep one eye on the forecasters' live models as results land: their credibility, like the pollsters', is being graded in real time.

By Wednesday morning, America will know who controls Congress. It will also know whether the industry that claimed to see it coming actually did.