Methodology
How Who Shows Up works, and how well it would have done
The short version: estimate the national mood, carry it down to every race, update each race with its polls, and simulate the whole election 20,000 times with errors that move together, the way real polling misses do. Every step was tested against elections the model had not seen.
1. The national mood
The model first estimates the national House popular vote from three independent readings, each corrected for how it has historically missed and weighted by how accurate it has been:
- Special elections. Our main turnout signal. In low-turnout specials, the party whose voters are more motivated beats the district's usual lean. We add the cycle's average overperformance to the last presidential margin. Historically this overstates Democrats by about three points, so the model subtracts that.
- The generic ballot. Polls of which party voters want in Congress, read in two-party terms. In late September these polls have overstated Democrats in 14 of the last 15 elections, by 3.4 points on average, and the model corrects for it.
- Fundamentals. The president's approval rating and the midterm penalty, fit on every election since 1978.
Each reading is a small Bayesian model: it learns from history not just a best estimate of its bias but a range of plausible values, so its uncertainty flows into the forecast.
2. From the nation to each race
A district's starting point is its 2024 presidential result under the new 2026 lines. The national shift is applied two ways, and each simulation blends them:
- Turnout-shaped: each party's vote in the district rises or falls in proportion to its vote nationally, the way an enthusiasm gap works.
- Persuasion-shaped: every district's margin moves by the same amount, the way voters changing sides works.
Across eight past elections, the turnout-shaped version explained about 65% of the pattern on average, but it swung from 0% to 100% depending on the year. So each simulation draws its own mix, and our uncertainty about what kind of year 2026 will be flows into the odds.
Then come adjustments measured from past races:
- Personal vote: incumbents keep part of how far they outran or trailed expectations last time: about half for House members, about a third for senators and nearly all of it for governors, learned from races since 2018 (senators' carryover has shrunk: it was about two thirds in 1996 to 2004). A House member's first re-election gets about 3.6 points more. That is why Vermont's Phil Scott, a Republican, is favored in a very Democratic state, and why a Republican who trailed Trump in their own district in 2024 is more exposed than one who ran ahead of him. Where we can't find the incumbent's last race (for example, they ran unopposed), they get the average House incumbency edge of about 2.3 points.
- Candidate experience: each nominee is coded from their record (senator or governor; U.S. representative, statewide office or mayor; state legislator or local office; none). We coded every nominee back to 2018 and tested what the gap is worth. In governors' races it matters when the race is close, up to about eight points. In House and Senate races it added nothing once the district's lean, the national mood, incumbency and campaign money were known: experienced challengers tend to run in the years and places that already favor them, and they raise the money to show for it.
- Campaign money: each side's money as of its latest report to the Federal Election Commission: cash in the bank plus everything spent this year. It counts only in close races, where a big fundraising edge is worth up to about five points in the House and more in the Senate (where polls usually outweigh it). In safe seats it counts for nothing, because incumbents there pile up money regardless. Adding it improved our backtest's accuracy, most of all in Senate races. Governors' races are not in FEC data.
- Candidate ideology (House): how far each nominee sits from the middle of their own party, measured from who donates to them (Adam Bonica's DIME scores, which run through 2024, so first-time candidates have none and count as typical for their party). In a close House race, the nominee further out from the center pays for it: a typical gap is worth about a point. It made no reliable difference in Senate races.
- A close-seat effect: in all four test elections, Democrats ran about one and a half points ahead of expectations in competitive House districts, even after accounting for money, so the model expects that.
3. Polls
Polls are collected every morning from three sources (VoteHub, Wikipedia and Decision Desk HQ), and a poll listed in more than one counts once. When a pollster releases several versions of the same poll, such as likely voters and registered voters, only one counts, preferring likely voters. Each race's polls are averaged in two-party terms. A poll's weight halves every 14 days (the best-performing half-life when we tested 3 to 60 days on past elections). Larger samples and likely-voter polls count a little more. Polls sponsored by a campaign or party are corrected for their measured lean toward the sponsor, about three to five points, and count half as much. That lean was measured on FiveThirtyEight's archive, so a polling firm counts as a party's side the way the archive treated it: firms it nearly always marked partisan, such as Trafalgar, are corrected, while firms it didn't, such as InsiderAdvantage, are left to their own track record below unless a campaign or party paid for the poll. We leave out Rasmussen Reports entirely, in 2026 and in the past years the model learns from: FiveThirtyEight dropped it in 2024 for not meeting its standards of objectivity and methodology, and leaked emails later showed it privately sharing results with the Trump campaign. Every correction here assumes we know whose side a poll is on.
Pollsters' own leans. Some polling firms run a little too Democratic or too Republican year after year. We measured each firm's record against actual results in more than 10,000 past race polls (after the corrections above, and relative to how all polls missed that year) and correct its 2026 polls by that amount. A firm with only a few past polls gets a correction pulled most of the way back to zero, and a firm with no record gets none. For well-known pollsters the corrections are small, typically under two points. The corrections adjust firms relative to one another but leave the average of all 2026 polls where it is: this year's busiest pollsters happen to be firms that have run a little Republican of the pack, and shifting every poll toward Democrats on that basis alone would be a bet the historical record can't test. It made the backtest's poll averages more accurate, especially on the eve of the election, when polls matter most.
Undecided voters. Reading a poll in two-party terms quietly assumes the undecided voters will split the same way as everyone else. In more than 11,000 race polls from 1998 to 2024, they didn't. Undecided voters split much closer to evenly than the decided voters around them, so a race polling 46-40 is less safe than one polling 53-47, and the model treats it that way. They also move with the national mood: races with more undecided voters than usual swung toward the party out of the White House in midterms, hard when a Democrat was president (1998, 2010, 2014 and 2022) and only slightly when a Republican was (2002, 2006; 2018 leaned slightly Republican). The model applies that pattern, and it widens the range for races with many undecided voters, since those have been harder to call.
How much a race's polls move it away from its fundamentals depends on how accurate poll averages have historically been at this point in the race. That depends on how many polls there are, and it improves as Election Day nears. A Senate race with 40 polls leans mostly on them. A House district with none runs on fundamentals alone.
4. Simulation
Each of the 20,000 simulated elections draws a national environment, a turnout/persuasion mix, a shared error for each state (a polling miss in Wisconsin hits every Wisconsin race), a shared error for each region, and each race's own error. That correlation is why the seat ranges are realistic rather than falsely narrow. Each race's own error has "fat tails": in past elections most races landed closer to the forecast than a bell curve would suggest, and a few missed by much more, so the model draws its errors that way too. That made the backtest's ranges more honest (its 80% ranges held 85% of results instead of 88%) and its probabilities slightly more accurate.
The same simulations find each chamber's tipping point. In every simulation, line up the seats the winning party took from its safest to its closest; the one that gave it a majority (the 218th House seat; the 51st Senate seat for Democrats, the 50th for Republicans with the vice president's tiebreak) is that simulation's tipping point. A race's tipping-point chance is how often it plays that role. Vote's sway divides that chance by the votes we expect in the race (its 2024 presidential vote, scaled by the state's 2022 midterm drop-off) and compares it with the average voter's, so it shows where a single vote is likeliest to decide control. Neither number changes any forecast; they summarize what the simulations already say.
What the model does not use
No pundit ratings (Cook, Sabato, Inside Elections), no other forecasters' models, and no prediction markets. PredictIt's prices appear on the Compare page and race pages for comparison, never as an input.
Track record
We rebuilt every election from 2010 to 2024 exactly as it looked on September 22 of that year, using only what was known then, and ran the same model on every House, Senate and governor race. Each year's national estimate and other fitted pieces were learned without that year. The headline numbers below are 2018 through 2024, the years the model can run with every ingredient; race polls for those years come from FiveThirtyEight's complete poll lists, counting only polls finished by that date and weighing them exactly as the live forecast does.
For 2010 through 2016 the model runs without campaign money, candidate records, ideology scores or special elections (none are coded that far back), and with only the last two months of polls. It still called
The last number is a simple yardstick: each race called from its district's or state's most recent presidential result plus the model's national estimate for that year, with no polls, incumbents or anything else. Lower Brier scores are better, and 0.25 is a coin flip.
When we said X%, did it happen X% of the time?
Each circle groups races by their forecast probability; its size reflects how many races are in the group. Points on the diagonal mean the probabilities meant what they said.
Chamber totals: forecast versus result
Shaded bars: the 80% range forecast on September 22. Tick: the forecast median. Dot: what happened. All 12 results landed inside the range. (Some districts are left out in redistricting years; the House rows count only the districts modeled.)
The national environment, 1996–2024
Each year predicted without that year in the training data, as of September 22. Shaded: 80% range. Small dot: forecast. Large dot: result.
Ideas we tested and dropped
Good models are defined as much by what they leave out. We built a county-level measure of how strongly each place's turnout amplifies national waves, a natural fit for a turnout-first model. On 2018 and 2022 results it did not predict how statewide swings distributed across counties any better than simpler methods, so it is not in the model. The broader turnout idea, shifting each party's vote in proportion rather than moving every margin equally, did pass, and it is.
We also tested whether kitchen-table conditions (consumer sentiment, how people rate their own finances, gas prices, inflation and unemployment) change how undecided voters break. They added nothing beyond the pattern above: hard economic times are already visible in the polls and special elections that set the national mood. Nor did turnout surges make polls miss by more; the biggest recent misses, in 2016 and 2020, hit every race alike rather than the ones with many undecided voters.
A large voter survey (the Cooperative Election Study, which checks official records to see who actually voted) showed that undecided voters turn out at about half the rate of decided ones, and that young and Hispanic undecided voters lean much more Democratic than others. So we tested whether undecided voters in states with more Hispanic or Black voters break more Democratic. In past polls they did, but adding it made the full model's forecasts of 2018 through 2022 slightly worse, so it is not in the model.
Primary turnout seemed like a natural enthusiasm gauge: if one party's voters flood its primaries, maybe they show up in November too. Using the Federal Election Commission's official results for every House primary from 2006 to 2022, states where Democrats' share of primary votes rose did tend to swing toward Democrats in the fall, but weakly, and the model already catches those swings through polls and special elections. Adding primary turnout didn't sharpen its calls in close races in any year, so it is not in the model.
Special elections set the national mood in this model, so we asked whether they also carry local news: if Democrats keep overperforming in one state's legislative specials, will that state's House races lean their way too? Averaged across a state's districts, sometimes. But in the close races that decide the House, using each state's specials made the forecasts worse in 2018 and 2020 and better only in 2022, so specials stay a national signal.
The model gives incumbents credit for how far they ran ahead of their state last time. We tested doing the same for challengers who have run statewide before, for Senate, governor or a state's single at-large House seat: Mary Peltola in Alaska, Sherrod Brown in Ohio, Roy Cooper in North Carolina. Across 91 such candidates since 1996, about a third of a past edge carried over, and it made forecasts from 2010 to 2016 more accurate. But it made 2018 to 2024 slightly worse, and recent elections are the ones we trust most: as politics has become more nationalized, candidates' personal appeal counts for less than it used to, and what remains already shows up in the polls. So the model leaves it out.
How this was built
Who Shows Up was built by Kurt Wirth, Ph.D., working with Claude, Anthropic's AI model. Claude wrote the code (the data pipelines, the model and this website) and explained the options along the way; Kurt decided what the model should do, which ideas to test and which to keep, and reviewed the results. No part of the forecast is generated by AI at run time: every number comes from the data and code described on this page, which are public on GitHub, and every ingredient had to prove itself on past elections before it was kept.
Sources
Polls from VoteHub (CC BY 4.0), Decision Desk HQ and Wikipedia. Election results from the MIT Election Data + Science Lab. District-level presidential results and special elections from The Downballot. Historical polls from FiveThirtyEight's public archive and Wikipedia. Past governor results from Wikipedia. Primary results from the Federal Election Commission. Presidential approval history from the American Presidency Project. Eligible-voter counts from the Census Bureau. The code is public on GitHub.