An Election Day Special With Nate Silver

Nate Silver · Silver BulletinNeutral

Prediction markets mostly mirror polls plus sentiment, lean slightly Republican, and can be moved by his model.

Silver says prediction markets reflect polling plus sentiment, which he rates as somewhere between smart and midwit. Forcing traders to quantify a price is useful, but the markets have a slightly Republican-leaning clientele and are sometimes driven by whales. Markets also react faster than polls to news such as scandals. He says his model's launch on August 11 moved the markets by several points, and that the markets were overconfident about Clinton in 2016. As of recording, his state-level probabilities were about 0.995 correlated with Polymarket's, though his overall Democratic win probability was a bit higher. He agrees with Polymarket on about half of the gaps between the two.

“I think they're reflecting polling plus sentiment in kind of a... It's somewhere between smart and midwit.”at 41:26

Nate Silver · Silver Bulletin

Polling errors have underestimated Trump, not Democrats, and the miss appears only when Trump is on the ballot.

He disputes the common view that pollsters systematically overestimate Democrats. He cites systematic underestimates of Trump in 2016, 2020 and 2024, smaller but directionally the same. He says that in every Trump-era election where Trump is not on the ballot, that error has not appeared. In New Jersey and Virginia last year, Democrats in governor's races were substantially underestimated. He allows that the pattern could be noise, and that Trump may be attracting low-turnout voters who pollsters struggle to reach.

“we have had systematic polling errors that underestimated Trump in 2016, 2020, 2024. A little bit more modest, but still directionally the same, right? In every election in the Trump era where Trump is not on the ballot, you have not had that error.”at 7:48

Nate Silver · Silver Bulletin

Democrats are favored to win the Senate, with prediction markets and his model both pointing to high-70s odds.

He pushes back on calling the Senate a toss-up. Polymarket and Kalshi had Democrats at about 68%, and his model had them in the high 70s. The case rests on Trump's approval at about 37%, the worst midterm approval for a president since Truman by some measures, a generic ballot lead of about nine points for Democrats, early voting and turnout favoring Democrats, and Republican candidates in red states like Texas and South Carolina being competitive. He also argues that Trump endorsing Paxton and going to war with Iran has cost Republicans. The House is also likely to flip, though he treats that as less of a sure thing.

“I would push back on the Senate being a toss-up. I think at Polymarket and [Kalshi], Democrats are up to like 68% or something as when we're recording this. And in our model, they're at 70... High seventies now.”at 12:53

Nate Silver · Silver Bulletin

Florida's DeSantis map could backfire on Republicans if Democrats win Florida by about nine points.

The map Ron DeSantis drew assumed a presidential-year Florida baseline. It was not built for a national environment about nine points bluer, in which Florida could move from R+13 to around R+4. If Democrats run the table in Florida, they could win more seats than under the old map. He thinks Texas Republicans are counting on Hispanic shifts in South Texas, which could backfire too. He says the House is not a sure thing, with about a 5% chance of a Republican hold, and that gerrymanders become more effective if polls underestimate Democrats.

“In Florida, Ron DeSantis drew a map for a presidential year baseline... It's not robust for a D plus nine climate where Florida goes from R plus 13 to R plus four or something. It wasn't designed for that.”at 27:48

Nate Silver · Silver Bulletin

Governor's races are underrated, and Democrats could end up governing almost two-thirds of the country.

He says governors' races get too little attention. Unlike Senate races, where Democratic leads are one or two points, many gubernatorial races are close. Per his modeling, Democrats could end up representing almost two-thirds of the country. He also flags Texas, where Abbott leads only narrowly, as important because Abbott has been the most aggressively pro-data-center governor.

“governors' races don't get enough attention... According to our modeling, we could wind up with almost two thirds of the country being represented by a Democrat.”at 29:36

Nate Silver · Silver Bulletin

Kansas is becoming more elastic and Democratic-leaning, so Hamilton's Senate challenge to Marshall is credible.

Silver argues Kansas has gotten more Democratic over time. It voted Democratic on an abortion referendum and has twice elected Democratic Governor Laura Kelly. He calls it a relatively elastic, well-educated state where opinions can change. He cites Senator Roger Marshall's unflattering medical billing practices and says Hamilton's pitch on cheaper gas and integrity could catch on. A recent CNN poll had Hamilton up four points. Silver says he has more trouble believing the South Carolina polling than the Kansas polling, which leads him to take the Kansas race seriously.

“Kansas has low key gotten ... a bit more democratic. It voted Democratic on this abortion referendum. It has elected Laura Kelly as governor twice.”at 38:22

Nate Silver · Silver Bulletin

Election models keep underestimating Republicans in Florida because net domestic migration adds red voters.

Silver says models should account for net domestic migration. Florida's models keep missing low because a steady flow of Republican-leaning voters moves into the state. The models use benchmarks from the last election, but Florida is a moving target for that. He ties this to the Senate race, where Democrat Angie Nixon is polling poorly and Florida shows no Democratic bump in the markets. The argument is that the state's partisan baseline is shifting, so older benchmarks understate the Republican vote.

“One thing we should probably account for in our models is net domestic migration. The models keep missing low on Florida because you have a steady flow of red voters into the state.”at 40:45

Nate Silver · Silver Bulletin

Consumer sentiment surveys are unreliable, with the Michigan survey's online shift oversampling Democrats.

He addresses the gap between very poor consumer sentiment readings and solid consumer spending and economic data. He argues that some consumer confidence polling is not very good. He says the University of Michigan survey moved to online sampling in a way that gives it far too many Democrats. He contrasts this with election polls, which face a reality check every two years and therefore have stronger incentives to get it right.

“I think some of the consumer confidence polling is not very good... clearly, I think the University of Michigan survey changed the composition to online in a way that just has, like, way too many Democrats in the sample.”at 20:20

Nate Silver · Silver Bulletin

AI is the one policy area where federal action could come before 2028 if timelines prove less aggressive than labs predict.

He expects a Democratic chamber of Congress to produce gridlock. Even so, he thinks AI is the one policy area where action could come before 2028. The reasoning is conditional: if AI timelines are even a third as aggressive as the labs claim, lawmakers will be forced to act. He also says he will look for evidence that AI was a campaign issue in these races and that candidates seen as too pro-AI were punished for it.

“I think the one policy area where you could see action before '28 is on AI. Mean, the timelines are even a third as aggressive as the labs think, then they'll kind of be forced to take action.”at 30:17

Nate Silver · Silver Bulletin

AI coding tools cut his model-building work sharply, making him far more productive, though they need guidance.

Asked whether AI has made his modeling work easier, Silver says yes. He calls the chatbots a great teammate. Coding the NFL playoff tiebreakers took him two and a half days last year, and now GPT or Claude does it in about fifteen minutes. He says code is roughly 10 times faster. He adds caveats: the tools can introduce bugs and hide them, and about two-thirds of his prompts still need substantial guidance. His view is that the value comes from pairing the tools with domain expertise, since coding was always his weak point and these tools are strongest exactly there.

“Last year when I built our model, it took me two and a half days to code them up... Now you have GPT or Claude do it, and they do it in fifteen minutes while you're at the gym... Code is 10 times faster.”at 46:46