An Election Day Special With Nate Silver

Nate Silver put Trump’s approval at about 37%, with only 19% strongly approving; by some measures, that was the lowest approval for a president at a midterm since Harry Truman in 1946. Republicans lost 41 House seats in 2018, when Trump was somewhat less unpopular. Silver said the political conditions strongly favored Democrats: midterm history, economic anxiety, the Iran war, Trump’s lame-duck status, strong Democratic early voting, and Democratic gains in New Jersey and Virginia in 2025. He expected Democrats to probably win the House and gain some governorships and state legislatures.

Silver’s model put Democrats’ chance of winning the Senate in the high 70s, compared with about 68% in prediction markets. He said the contest was difficult to model and calling it a toss-up would not be honest: Democrats would have to win in very Republican states such as Texas, Kansas, and Alaska, and possibly South Carolina. Without Trump’s war with Iran, Silver would have viewed the Senate as a toss-up rather than leaning Democratic. He said a Democratic Senate loss would be a very bad night for pollsters.

Democrats led by about nine points on the generic ballot. If polls were accurate, Silver said, that would create too much of a headwind for Republicans and could give Democrats a big night. But he cautioned that polling errors can be correlated across similar states, so individual races should not be treated as independent outcomes. Polls underestimated Trump in 2016, 2020, and 2024, though the 2024 error was more modest; the pattern had not appeared in Trump-era elections when Trump was not on the ballot. Silver floated the possibility that Trump attracts low-turnout voters who are hard to reach or screened out by likely-voter models, while stressing that this remains a theory. Errors are less correlated across states in midterms than in presidential elections because the candidates differ.

Early-voting indicators appeared favorable to Democrats, including comparatively strong participation by Black and low-propensity voters, with Democrats gaining more low-propensity voters than Republicans. Silver warned that he is not an early-voting expert and that analyses can be ad hoc, without clear rules. Differential turnout matters more in midterms and can help produce large waves. He said South Carolina, Louisiana, and Mississippi have low elasticity, meaning their results are relatively insensitive to national shifts. In an extreme South Carolina simulation, Black turnout at 2024 rates and white turnout at 2022 rates moved the topline by about 10 points; differential turnout might be needed for Democrats to flip the state. His model gave Democrats a lower chance in Mississippi than Polymarket’s 14% estimate, partly because of that low elasticity.

Silver said Trump remains central to the election even as many Republican candidates have stopped featuring him in ads and Republican podcasts. He argued that Trump’s prominence works against Republicans’ apparent incentive to let more moderate candidates distance themselves. Trump’s endorsement of Ken Paxton may have helped give James Talarico a slight lead in Texas. Texas Republicans’ map gains could depend on South Texas Hispanics continuing to vote Republican rather than behaving as swing voters, leaving room for Democratic incumbents to hold on. Mapmakers must take risks to maximize seats, and a roughly 5% scenario with a less favorable polling environment could make Republican gerrymanders more effective. Florida’s map, drawn around a presidential-year baseline, could give Democrats more seats than the old map if they run the table.

Silver said his Flipper model aggregates polls and translates them into probabilities, while prediction markets combine polling with sentiment and require traders to quantify odds. At recording time, his state-level probabilities and Polymarket were about 0.995 correlated, though his overall Democratic win probability was somewhat higher. He agreed with Polymarket in about half of the races where they differed, and said market participants may lean slightly Republican. He rejected the simplistic view that polls always underestimate Republicans, noting that markets were more overconfident than his model about Hillary Clinton in 2016. He also criticized the University of Michigan consumer-confidence survey’s shift to online sampling, saying it left too many Democrats in the sample, and warned that AI-generated respondents cannot replace original polling data or capture meaningful variation within broad groups such as Hispanic voters.