As of 5 November, when this was written, Scott Colom leads Mississippi's Senate race by 1,724 votes with about 96% counted. We had given him a 4% chance.

There is a river between what polls can see and the vote on the other side. A prediction has to cross it. Mississippi was our hardest crossing. It had three public polls all year, all sponsored by Democratic-aligned groups, and not one independent public poll. Our model discounted all three, and hardly anyone questioned it. Come on! It is Mississippi, a Black Democratic candidate, in a state as red as blood. So no one cared to look closer. Nevertheless, the results came as a big shock, or a slap in the face, depending on who you ask.

Now, how well did we do in predicting this election cycle?

Three verdicts

A prediction answers three questions. Who wins? By how much? And how wide a margin of error should we allow before trusting it? We answered the first very well. We answered the second poorly. Our answer to the third was too narrow on the Democratic side.

We gave Democrats a 94.6% chance of House control and a 77.8% chance of Senate control. They won both. Our favourite won in 475 of 506 races, counting current leaders in uncalled races. We also checked every probability against the outcome with a standard measure called the Brier score, which punishes confident misses most. A perfect record scores zero, and calling every race a coin flip scores 0.25. We scored 0.049. Put simply, our misses added up to about a fifth of what pure guessing would have cost. Many of the 506 races were safe seats, which any model gets right, so the close races remain the real test.

Of the 31 races where our favourite did not win or does not currently lead, 30 went to the Democrat. The exception was Vermont's governor race, where Republican Phil Scott won by seven points after polls had the Democrat ahead.

Every result is on the Democratic side of our median. The House passed the top of our 80% range. We gave 262 or more seats a 6.8% chance. The Senate could finish at the top of its range too. Governors are in the upper half. The counted House vote is inside its range, though projections that assume late ballots lean Democratic put the eventual margin nearer D+13.

The one-sided pattern also shows up race by race. Of ten races where we gave the Democrat a 30–40% chance, Democrats won eight. We saw the direction of the election. We did not measure its distance.

In other words, we did see a blue wave coming, but no one had the guts to call it what it really was, a tsunami!

Where we lost the crossing

Mississippi had one public poll in April, one in August and one in late October, roughly four months apart. They showed Republican Cindy Hyde-Smith ahead by three, six and three points. Our final prediction put her ahead by about ten. We rated the race Likely Republican.

Polls paid for by a party tend to flatter their own side. Our model therefore moves Democratic-sponsored polls 3.8 points toward Republicans, moves Republican-sponsored polls 3.1 points the other way, and counts each as about a third of a normal poll. These rules come from 20,000 past polls. All three Mississippi polls were Democratic-sponsored, so together they counted for less than a quarter of one poll. For races with no polls, the model borrows the polling errors of similar states. Mississippi had polls, so this backup never switched on.

Fundamentals and expert ratings pulled the prediction further toward Hyde-Smith. Inside Elections rated the race Safe Republican. Our Mississippi blend was about 40% polls, 44% fundamentals and 17% expert ratings. The historical expectation won inside our model. The thin polling signal came closer to the count.

That choice was ours. Sponsor discounts protect against a real risk, but this time the discount hurt. Our safeguards against a misleading poll could not tell us when the few polls we had were useful.

The problem extended beyond Mississippi. In polled races with at least 95% of the expected vote counted, raw poll averages beat our blended prediction in Senate and governor races.

Polled races	Raw polls	Our prediction	Fundamentals	Expert ratings
Senate (23)	+2.8 (4.7)	+4.1 (5.3)	+6.7 (8.2)	+10.2 (10.5)
Governor (20)	+1.5 (3.7)	+2.5 (4.3)	+2.2 (7.2)	+6.3 (8.8)
House (82)	+4.5 (6.3)	+4.8 (5.8)	+5.7 (6.3)	—

Positive numbers mean an average underestimate of the Democratic candidate, in points. Mean absolute error is in parentheses.

Expert ratings were the furthest off in the Senate and governor samples. We added the experts' view to keep the polls from leaning one way. It ended up working against our effort to get a real picture of how the electorate behaved.

Our inference is that ratings can re-read polls through an expectation that historically Republican states will stay Republican and historically Democratic states will stay Democratic. That expectation can help where polling is thin. This year it pulled against the result. In House races, where district polls are sparse and noisy, our blend still reduced absolute error.

We also subtracted 2.7 points from the national generic-ballot average before using it. The correction was fitted on earlier cycles. This year, the counted national vote is more Democratic than the poll average. By our rough estimate, removing the correction would have raised our expected House total by about five seats. That is roughly a quarter of the gap between our expected total and the current result. Most of the gap remains.

Pollsters have changed their methods too. Nate Cohn reported before the election that, under its old weighting, Times/Siena's final polls would have been about three points more Democratic in seven states. If other firms made similar changes, our historical correction may have applied a comparable adjustment twice. That is a possibility, not a finding. The industry's reviews are still to come.

All this is to say we now have another way of correcting ourselves. The view that polls always underestimate Republicans now has a rival. Polls can underestimate Democrats too. Or simply get "corrected" a bit too much!

The race we could see

South Carolina had 14 polls, averaging a Republican lead of 0.7 points. We gave Democrat Annie Andrews a 35% chance against Republican Darline Graham, who was appointed after Lindsey Graham died in July. Andrews won by about 4.5 points.

Our archive of 508 unique headlines from 13 national outlets named South Carolina only three times between 12 September and 3 November. Texas appeared in 23 headlines and Alaska in 20. South Carolina was one of the races that made our 22% chance of Republicans holding the Senate plausible, yet it barely entered the conversation.

Why is that? Why couldn't anyone get South Carolina into the conversation? The fact is that the United States has a big and complicated map, and people can only focus on so much. So they pick four or five races, and by the time they are done discussing them, everyone is too tired.

This election, it was Michigan, Texas, Ohio, Iowa and Kansas. If you were really invested, you might go as far as Alaska, and maybe mention Georgia. But talk about South Carolina? You've got to be joking. Mississippi? You don't even know this country.

There were signs of a contest in South Carolina, though. A Republican Senate group committed $5 million on 6 October. A late-October NYT/Siena poll was tied. Those signals said "competitive," not "Andrews will win." Our 35% said the same thing.

We missed these signals, or could not include them in our models, because our methods were not built to capture them.

The current changes

Single-race polling misses of three or four points are ordinary. What stood out here was their shared direction. The river moves between cycles. In 2016 and 2020, polls overstated Democrats. The 2020 error was largest in heavily Republican, whiter, less dense states. In 2024, AAPOR found polls overstated Democrats by 2.7 points on average. Then, by Silver Bulletin's measure, polls in the 2025 New Jersey and Virginia governor races and two House special elections underestimated Democrats by 6.1 points. Our 2026 comparisons show another Democratic underestimate, with some of the largest misses in red Southern states. The geography looked familiar. The direction reversed.

We will fit our national bias correction on midterms, give recent cycles more weight and account for pollsters' changing methods. When the correction's direction is uncertain, we will widen the range rather than shift the centre.

We will re-test our fundamentals and expert weights out of sample. We will give races with tiny effective polling samples a "thin polling" category that borrows information about polling misses from similar, better-polled states. Several other Southern Senate races, including Texas, Georgia and South Carolina, missed toward Democrats by five to nine points. Our Mississippi prediction had no way to learn from them. We will flag polling deserts, where every poll comes from one side's sponsors, and test a regional error that scales with state partisanship. Each change will be backtested on elections from 2018 through 2026 before it ships.

We will test campaign behaviour, markets where polls are missing, special elections, citizen forecasts, and local news, podcasts and YouTube, under rules written before an election. Markets had Colom at about nine cents in October, so they offer no easy Mississippi fix. Social-media attention has no reliable record of predicting elections in advance. A vivid late signal is easy to admire after a surprise. It earns a place in a prediction only by working in advance. We will publish those test results whether they help or not.

We hope pollsters weight for the electorate expected this cycle, publish response rates by party as Times/Siena did, and run independent polls in states that attract little attention. One nonpartisan Mississippi poll in October would have changed our range. Panels about what voters care about and how they get their news are worth testing too.

All to say, polls are a great tool. But they need a little more help to get the picture right. We have found a few ways to correct for them, and there is still a lot to learn and adjust. Before we cross the wide Mississippi river again, we need to check where the current is running.