Can we still trust the polls & forecasts in 2026?
Democrats have been traumatized ever since they witnessed some disappointing presidential election nights after being favored in the polls. Polls overstated Democrats in 2016, 2020, and, to some extent, in 2024. In these cases, national polls were actually quite accurate, but states decide the Electoral College, and state polls overstated Democratic support quite badly in some key swing states.
That poses an interesting question for the upcoming midterms. The congressional generic ballot has been accurate in past cycles, especially in midterms. The Senate, however, is decided by statewide races, so we have to rely on state polls & forecasts. In the past three presidential cycles, Senate polls badly overestimated Democrats, but luckily for them, we are only in a midterm cycle, and Trump is not on the ballot. Do Democrats still need to be afraid?
I took a look at the 2022 and 2018 midterm Senate forecasts by FiveThirtyEight’s Deluxe Model, the former gold standard of election forecasts, to see which party got overstates where. Looking at the forecast misses in each race, we get two very different pictures. Depending on which polling miss you believe to be the more likely scenario for 2026, you’ll get a drastically different prediction.
Republicans Underestimated in Most, but not All Key Senate Races in 2018
The 2018 midterms are the current go-to election for comparisons because they happened during Trump’s first term. While he wasn’t on the ballot, it was, in part, a referendum on his presidency similar to the upcoming 2026 midterms.
538’s generic congressional ballot forecast for the US House was highly accurate. Democrats won by 8.6 after being forecasted to win by 8.8%.
The Senate is a different story. Republicans were favored to keep control of the chamber as Democrats faced a brutal map, but 538 forecasted Republicans to only gain between 0-1 Seat. In the end, they gained two. In the Senate, two seats go a long way. That miss is largely explained by forecast misses in a few key races, where incumbent Democrats had to defend states Trump won in 2016.
The two key misses here that led to Republicans’ net gain of two seats are Indiana and Florida. Indiana was classified as a toss-up leaning towards Democrats by 1.9 points, but Republicans won the state by almost six points. Florida was categorized as lean D by a margin of 3.4 in favor of incumbent Democrat Bill Nelson. The race became the closest in the nation and went to Republican Phil Scott by just 0.2 points. That 3.6-point miss is not bad at all for a forecast, but it makes all the difference in close races.
These were the two deciding misses, but 538’s forecast also underestimated Republicans in other battlegrounds. The forecasters expected Democrats to lose their Senate seats in deep-red Missouri and North Dakota, but they overstated Democratic support by 5-6 points. Ohio and Michigan, two Midwestern states, where Trump outperformed polling expectations, saw Republicans do 6-8 points better than forecasted. The only bright lights for Democrats in this comparison came from Nevada and Texas, where their candidates did almost 3 pts better than expected.
Other Key races were fairly accurate. Four of the six closest key races had acceptable forecast misses below 3 pts (Arizona, Texas, Nevada, Montana), and the other two were within a decent range, with misses between three and four points (Florida and West Virginia).
Across the 10 key races I identified – North Dakota through Nevada – 538 underestimated Republicans by 2.79 points. Across all races, they were underestimated by 2.76 points. So the error stretched across the map. That is a key difference from 2022.
Democrats Underestimated in ALL Key Senate Races in 2022
2022 paints a completely different picture. It was the only time since probably 2012 that Democrats outdid polling expectations. In the House of Representatives, Republicans were huge favorites, but they ended up with one of the slimmest majorities ever.
In the Senate, Republicans were slightly favored to win back control as they only needed to flip a single seat. The morning after Election Day, it became clear that Democrats had defended all their vulnerable incumbents and even picked up an open seat in Pennsylvania.
They outdid 538’s forecasted margin in all key races.
They central finding here is probably that the forecast, and therefore the underlying polls, systemically underestimated Democrats in the key contests, but no so much in the rest of the races.
In the 10 key races I identified1, Democrats outperformed 538’s forecast by an average of 3.31 points. That’s about half a point more of an overperformance than Republicans had in 2018 in key races, but there is an even more interesting finding. Republicans’ 2.79-point overperformance in key races was almost identical to their overperformance across all races in 2018, meaning they were underestimated across the map. That uniform forecast miss points to an issue with fundamentals, rather than battleground polling.
In 2022, however, Democrats were only underestimated by 0.4 points across all Senate contests. So the systemic forecast miss was almost entirely confined to key races. Yes, Democrats were largely underestimated in deep blue states, but so were Republicans in deep red states. The two practically cancel each other out.
I can only think of one reason for that discrepancy: polling. Key races saw countless election polls that heavily influenced 538’s forecast, while obscure contests were largely built on fundamentals. While the largest misses happened in these races, they were more random and not systemic. Races that were heavily polled had smaller misses, but all favored Democrats. My read: Late, Republican-associated polls (by the likes of Insider Advantage & Trafalgar) overstated Republican support and tilted the averages in their favor.
This is also the reason I choose to use 538’s forecasted margins. They did much better than the simple polling averages from Real Clear Politics, partly because 538 weighted pollsters more sophistically and used a mix of indicators in their deluxe model. RCP overstated Republican support by an even larger margin.
Summarizing Our Findings: What Does it Mean for 2026?
In 2022, polls in key races underestimated Democrats: While 538’s Senate forecast was very accurate overall, polls skewed the forecast in the top 10 battlegrounds in favor of Republicans by 3,31 points.
In 2018, 538 underestimated Republicans across the board (2.76 points), and in the 10 key battlegrounds (by 2.79 points). Given the evenly spread nature of that forecast miss, the reason might not just have been the polls but the fundamentals the team used back then.
As G. Elliot Morris pointed out: If you try to account for past polling biases in your forecast for 2026, you’ll end up doing worse. If you’re a Democrat, you are probably most worried about states like Ohio, where polls have overstated Democratic support in the past, and more relaxed about New Hampshire or Georgia.
The 10 key races of 2022: Ohio, North Carolina, Wisconsin, Nevada, Georgia, Pennsylvania, Arizona, New Hampshire, Colorado, Washington
It makes me wonder whether Cook, Sabato et al. May be closer to the mark than the polling. G. Elliott Morris just had a piece up that indicated that the prediction models tended to give an advantage to the incumbent president’s party in mid-term elections relative to the polls 45 days out.
This is an interesting question and design. I like the tables. I apologize if I am just missing this, but I'm not able to unpack this paragraph, which is summarizing some important findings:
"Key races saw countless election polls that heavily influenced 538’s forecast, while obscure contests were largely built on fundamentals. While the largest misses happened in these races, they were more random and not systemic. Races that were heavily polled had smaller misses, but all favored Democrats. My read: Late, Republican-associated polls (by the likes of Insider Advantage & Trafalgar) overstated Republican support and tilted the averages in their favor."
In the third sentence, I think you mean heavily polled races in 2018 had small pro-Dem misses, notraces in 2018 and 2022 - is that right? Also, in the last sentence, which polling averages were tilted in favor of Rs by late, Republican-associated polls? Is that a 2022-only effect, or is there a more complicated interaction between other polls and R-affiliated polls?