“I’m part of the greatest band in the world. Am I happy with that? No I’m not. I want more!” – Liam Gallagher
At State Navigate, we celebrate our wins and learn from our losses. That attitude toward election forecasting (and now polling) has helped us prove remarkably accurate in informing the public about where state legislative elections are heading going into an election, and, of course, has helped us inch closer to perfection each cycle.
With most state legislative races now called, we can share our analysis of our performance. This year was our most successful forecasting year yet (including our work at our previous website, CNalysis), despite a large, industry-wide polling miss similar to the one in 2020, but in the opposite direction. We saw this error coming and made our peace with it as we adjusted our forecasts as best we could in most states.
Total Recall Bias, Our Environment Estimations, and the Polling Industry
Noah Rudnick flagged to us in late September that polling would probably miss again, as it did in 2025, underestimating Democrats due to conservative weighting decisions. Rudnick gave us this heads-up in 2025 and was right. We privately tracked Rudnick’s theory throughout October, creating histograms showing where recall measurements fell in every poll that published the number, and found that his hypothesis would likely prove correct in a few weeks. During this timeframe, we conducted our final polls in Pennsylvania and noticed a disproportionate rate of recall switching among Trump voters over a period when the president’s approval declined by just three points. As a result, our team started modeling and measuring the effect of “total recall bias” on the generic ballot, landing on two points and making recalled toplines appear bluer than they actually were. Polls missed by five to six points, though, and that’s where our ancillary analysis in the Total Recall Bias article that builds off Rudnick’s theory comes into play: it shows how polls weighted their electorates toward Republican-friendly years like 2022 and 2024, rather than a favorable Democratic environment like 2018.
Two weeks ahead of the election, as we looked at recall bias and continued to analyze the mistakes pollsters were making, I published an official estimate of the generic ballot landing between D+11 and D+14; the latest estimates show that the final results are close to the provided ceiling. Based on 2018 exit polling alone, I estimated the recalled vote in the electorate’s answers would land at Harris +4.5; the final exit-poll result was Harris +5. I admittedly hedged, though, and landed at Harris +3.75. With Total Recall Bias, our rough estimate of the actual electorate’s recalled vote is Harris +2.0 to 3.0, but we look forward to more rigorous modeling and analysis to determine the actual number from folks like David Shor, who estimated the 2022 electorate to be Harris +0.6.
Our determination of the environment was rather elementary: President Trump’s approval is several points worse than it was in 2018, so the environment should be at least slightly bluer than 2018, if not several points bluer. Most pollsters refused to follow the strongest predictor of electoral environments (presidential approval), resulting in what appears to be the largest self-enforced error in polling this century; 2020 was due to the pandemic, but this one rests on the shoulders of pollsters who worried about underestimating Republicans. This error drove most (though not all) of our overall misses this year: our forecasts’ national environment estimate was D+8.9, an estimate that was derived from giving 65% of the weight to the generic ballot average and only 35% to the approval-based estimate, a weighting scheme we plan on changing going forward due to the evident cognitive biases observed in pollsters’ weighting decisions this year.
These takeaways are not condemnations, but constructive criticism for an industry that we partly rely on to help our forecasts. We love the polling industry and want it to improve. We also needed polls in many states; there was oversaturation in places with hot-topic Senate races like Texas, Ohio, and Iowa, while important gubernatorial races in Tennessee, Oklahoma, and Nevada were ignored. Hopefully this doesn’t happen again, but if it does, we’ll be happy to fill the gaps by providing the public with statewide polling where needed, as long as we’re provided with the bare-minimum funding and resources to do so.
Our Forecasts’ Successes
This marked the first year that our forecasting team created mixed-method (quantitative + qualitative) forecasts for every single district, including the multi-member districts; in 2024, our expanded model (our mixed-method model that year) only applied to single-member districts. Our 2026 forecasts thus gave a prediction for nearly every state legislative seat in the country: only a handful of special elections we were unaware of, or that were in chambers not up in 2026, didn’t make the cut. As of this writing, we’ve identified the winners of 6,117 state legislative races, with about 70 seats still counting votes or heading to recounts.
When we exclude uncontested races, 93.7% of candidates our forecasts identified as favorites won. That’s about one point less than 2024, despite polling being five or six points less accurate in 2026: a remarkable achievement thanks to our backend work that involves my qualitative ratings. It’s also one point less than 2022, the most accurate year for polling in years when we’ve worked on a forecast. However, this comparison to past performances isn’t entirely apples-to-apples. In 2020 and 2022, we published qualitative ratings for every single-member state legislative district. In 2024, we published qualitative ratings for every type of state legislative district and used a mixed-methods model only for single-member districts. In 2026, we only published a mixed-methods model. When we isolate to the single-member districts, our forecast performed only 0.4% worse than in 2024. Multi-member districts are not only more difficult to forecast due to them being different types of elections, but they also are more often than not in very small districts, which are the hardest to forecast, especially without any on-the-ground guidance and conventional wisdom, which was our case this year in every state with multi-member-district chambers.
Some states were kinder than others to our forecasts, and our team knows why. Alaska, Wisconsin, Pennsylvania, and Michigan, states where we spent the most time keeping in touch with in-state experts to compare notes with, due to each having highly competitive chambers, were some of our best. In-state experts we kept in close contact with also helped us avoid catastrophe (a double-digit miss rate) in states like Kansas, Texas, North Carolina, South Carolina, Utah, and Georgia, softening the blow of polling misses there that hurt our forecasts. To our fortune, the West Coast was also remarkably kind to our forecasts. Other in-state experts we spoke with in New York, Colorado, Wyoming, Hawaii, Maryland, Nebraska, and Illinois also helped guide our forecasts, delivering them with a miss rate below our national average.
The districts that are always the hardest to forecast are in New England. These districts have very small populations and are in electorates where candidate quality and community connections are the strongest drivers of vote choice, despite reliably Democratic partisanship at the federal level. Our forecasts perform worse there than our national hit rate, which is no surprise: we lack connections with in-state experts in this region, something we hope to remedy for 2028.
Other states where our forecasts performed poorly were very conservative states, where the environment was far more Democratic-friendly than one would expect. This was exacerbated in Indiana, North Dakota, and Arkansas, where we not only failed to make the time to adjust our qualitative ratings accordingly for environmental expectations, but also lacked in-state connections.
Overall, though, our forecast’s probabilities were far more reflective of outcomes than in 2024 despite a slightly larger percentage of upsets in outcome. A forecast is calibrated when the things it calls 70% likely happen about 70% of the time. In the middle of the range, where races are actually competitive, our 2026 forecast was close to that. Seats we gave a 60 to 70% chance were won by the favorite 65% of the time, and 30 to 40% underdogs won 35% of the time, etc.
Our chamber forecasts are no different in their improvements. Going into Election Day, we gave odds of control for 88 state legislative chambers, and the favorite won 86 of them. Our own odds said to expect about four upsets; but there were two: the Texas House, where we gave Republicans a 73% chance of holding on, and the Georgia House, where we had Republicans at 65%. Going into the final month of the election, we were told by in-state sources that the Democratic odds in both chambers should be higher, and by the final week of the election I told our team to prepare to take a loss in the Georgia House due to very poor polling in the state. In the 15 chambers where neither party had better than a 90% chance, the favorite won 13, and the party that ended up in control had been given a 70% chance on average. That’s our best showing since we started forecasting chambers at CNalysis in 2020: our favorites won 92% of chambers in 2020, 94% in 2022 and 96% in 2024, and our Brier score, a standard measure of how closely probabilities match outcomes, has fallen every cycle, from 0.048 in 2020 to 0.024 this year.
As one would expect, the upsets were mostly driven by unexpected Democratic victories. Democrats beat our projected margin by 3.0 points on average. Democratic overperformances wasn’t the case everywhere, though, most notably Vermont, where Phil Scott (R) won by 3 points more than expected and voters clearly resented the Democratic candidates for state legislature at a far greater rate than our forecast expected when examining previous ticket-split rates between the gubernatorial and state legislative results. Not all politics is national: some states have overperformances that are likely correlated with the public opinion of their state governments and governors (e.g. Republican overperformances in Vermont, Oregon, Minnesota; larger Democratic overperformances in Indiana, West Virginia, Utah, Kentucky). Unfortunately, there’s a lack of consistent polling of voters’ attitudes on state government; the only frequent data point for measuring voters’ attitudes toward their Governor is from Morning Consult, a pollster that frequently publishes numbers we believe far too favorable toward incumbent Governors. If we had the resources, we’d love to conduct our own statewide polls of state governments everywhere in the future to give accurate reflections of voters’ attitude toward their Governor and state legislatures, since it is directly correlated with state legislative vote choice. Such data would further help our forecasts’ calibration.
Our Successes in Polling
Our polls fared better than our forecasts, which of course rely in part on other people’s polls. The same story this year rang true in 2025: had other polls just been closer to State Navigate, our forecasts would have been closer to the actual result, since our polls were.
State Navigate had the most accurate poll of the Georgia and Vermont Governor races, and was tied for third place in the Minnesota Governor race. Our Virginia poll was our most accurate statewide poll we’ve ever conducted, having just a 0.8% difference from the actual results statewide. That Virginia poll should have been the shot heard around the world for other pollsters given our known expertise in the Commonwealth– it implied a D+14 national environment this year, and that’s the result we got. Virginia usually trends in line with the nation in midterm elections, give or take a point, so betting against the Commonwealth’s electoral trends has once again been proven to be a mistake. Polls also bet against the actual 2025 results as well; our poll showed the electorate that year to be 6 points to the left of 2024 in recalled vote. The average poll in the lead up to the election, as we wrote, was only about 2 to 2.5 to the left.
We strive for perfection in our polling, though. We’re pretty happy with our performances in Vermont and Minnesota in terms of absolute error, but despite placing first amongst Georgia pollsters, we find our performance there to be milquetoast. In Pennsylvania and West Virginia, our polling was frankly near-perfect. That’s because we rely on party registration in these states rather than party ID for partisan weighting decisions.
Conclusion
Overall, this was an incredible year for State Navigate in the 2026 general election, especially when factoring in for the industry’s polling misses. We had our best-ever performance in predicting the winner of state legislative chambers, our forecasts’ calibration improved, and we entered the top 10 pollsters in the nation.
Like I said, though, we want more. We are perfectionists. We have some lessons from this year and some planned changes in methodology that should help, and we’re cautiously optimistic that polling firms will learn their lessons from this year and deliver us better averages to rely on. We’ll also work diligently on finding more in-state sources for us to communicate with to give our forecasts the on-the-ground perspectives that a regular quantitative forecast would fail to consider.
Cheers.
