An experiment with Claude to predict the outcome of all 435 House and 35 Senate races. Will run every Saturday from October 10 until Election Day.
F·R·O·N·T: Fundamentals, Ratings, Observed polls, National environment, Tails.
Loading the forecast…
Color reflects the modeled probability the Democrat or Republican wins. Hover for race details; click a state or district to pin a card with more information. Use the slider to adjust the map for a different national result.
The deterministic call is made for the candidate with a probability above 50%. Probabilities already include correlated national, regional and state error. The three rating columns are the main handicappers (Cook Political Report, Sabato's Crystal Ball, Inside Elections); an orange outline marks a rating a full step away from the others, and a "split" tag marks a gap of two steps or more. Polls shows the number of public polls in the window and their adjusted average. Click a row for the components and the race's history across runs.
| Race | Held by | Forecast | P(Dem) | Margin | Cook | Sabato | Inside | Polls · n · avg | Δ run |
|---|
Events inside the polls' lag window are applied to the national environment with decay; the playbook pre-registers how each class of surprise will be treated so the model can't be re-fitted after the fact.
F·R·O·N·T: Fundamentals, Ratings, Observed polls, National environment, Tails. Parameters come from the historical record, not from this cycle.
"Called" counts every race for whichever side is above 50%, so a 58% favourite counts as a whole seat. "Expected" adds up the probabilities, so that race counts as 0.58 of a seat. When most of the close races lean the same way the call runs ahead of the expectation: in the Senate this week Democrats are the slight favourite in Alaska, Texas, Maine and Ohio, which adds four whole seats to the call but only about 2.5 to the expectation, and the Republican-leaning toss-ups (Iowa, Kansas, Nebraska) pull the expectation down a little more. The call is the best single prediction of each race; the expectation and the range are the best description of the chamber.
The anchor is E, the expected national House popular-vote margin. It blends the generic-ballot averages (five aggregators, with the part of a lead above six points shrunk by 10%, following Moskowitz's 2004–20 slope of 0.87) with two weaker priors: a presidential-approval model fitted to 1994–2022 midterms (president's-party margin ≈ −18.2 + 0.29 × approval) and the special-election overperformance since January 2025, discounted because specials have overstated the eventual swing in every cycle since 2018. Polls carry about three-quarters of the weight now and more each week. The error on E is a Student-t (5 df) with a standard deviation that falls from 3.25 points today to 2.5 in the final week (the historical RMSE of the final generic-ballot average in midterms), plus a 3% mixture at twice that scale for a 2016/2020-style industry miss.
Fundamentals: partisan lean (twice the Cook PVI, or the 2024 presidential margin under the new lines in the nine redrawn states, relative to the nation) + E + incumbency (+3 House, +1.5 first-termers, ×0.7 if the seat was substantially redrawn; +2 Senate, +1 appointed) + a personal vote (40% of the incumbent's 2024 overperformance against the presidential vote, capped at ±5) + candidate quality (±1.3 per point of a −2…+2 experience/scandal score in the House, ±1.25 in the Senate) + a local economic-stress term from the state panel, capped at ±1 point against the president's party.
Ratings: Cook, Sabato, Inside Elections, Decision Desk HQ and Split Ticket, converted to margins (toss-up 0, tilt 4, lean 7, likely 12, solid 25; DDHQ probabilities via a 6.5-point normal) and averaged, with +0.5 toward the out-party for rater lag. Seats no rater lists get an implicit "solid" on the side the fundamentals favour, at lower weight.
Observed polls: district and state polls from the last 45 days, recency-weighted, with partisan internals shifted 3.5 points toward the other side and half-weighted. The poll weight rises with the number of independent polls (cap 50% House, 75% Senate). Where there are none, which is most of the map, fundamentals and ratings carry the race: lean, the national environment, local economics and candidate quality.
20,000 simulations. Each race's margin = expected margin + β × (national shock + regional shock σ1.5 + state shock σ1.5) + local shock. Local error is a Student-t (6 df) with σ 5.5 for well-polled districts, 6.5 with one poll, 7–8 with none; Senate 4.0 / 5.5 / 7.5. Senate β is a state elasticity (0.8 in the Deep South to 1.2 in Vermont). The scenario slider on the map fixes the national shock at a chosen value and re-scores every race with the same regional, state and local draws, so it shows which seats move together when the polls are wrong. Nebraska's independent sits in the Democratic slot but is counted separately; Alaska's independent is treated as the Democratic-aligned candidate. Chamber control: 218 House seats; 51 Senate seats (the Vice President breaks ties).
Markets are shown for comparison and get no weight: political contracts are under-confident and thinly traded at the race level, missed 2022, and misfired in several 2026 primaries.
| Parameter | Value | Evidence |
|---|---|---|
| σ national, now → final week | 3.25 → 2.5 (t, 5 df) | RMSE of generic-ballot averages vs result, 2006–2024: 2.75 at four weeks, 2.5 final (midterms) |
| National bias | 0 | Midterm misses alternate in sign (2006 +3.6, 2010 −2.6, 2014 +3.3, 2018 −1.1, 2022 +0.3); un-skewing worsened accuracy in 10 of 14 cycles |
| σ local, House | 5.5 / 6.5 / 7–8 | House district polls: ~6.2 average absolute error 1998–2018; unpolled seats ~9 (FiftyPlusOne) |
| σ local, Senate | 4.0 / 5.5 / 7.5 | Senate poll error 4.1 (2024) to 6.7 (2020); averaging and fundamentals reduce it |
| Incumbency | House +3 (+1.5 freshman); Senate +2 | 1.4–2.9 pts of vote share 2018–22; Split Ticket WAR 3.5 margin (2020) |
| Personal vote persistence | 40%, cap ±5 | Candidate effects persist ~50% cycle to cycle (Split Ticket); trimmed for overlap with the quality score |
| Approval → margin | −18.2 + 0.29 × approval | OLS on 1994–2022 midterms (residual SD 2.9); implies D+7.1–8.3 at 37–38% approval |
| Specials → margin | 0.6 × overperformance, ≤10% weight | Specials beat the realised swing every cycle since 2018 |
| Rating → margin | 0 / 4 / 7 / 12 / 25 | 538's average margins by rating (toss-up 0, tilt 4, lean 7, likely 12); Lean favourites win 77–94% |
| Local economy | ±1 pt cap | de Benedictis-Kessner & Warshaw 2020; effects small and run through the president's party |
By Jeff Reid ·
A Claude experiment: An attempt to forecast the 2026 midterm elections with a weekly call on every House and Senate race leading up to November 3, 2026. FRONT: fundamentals, expert ratings, observed polls and the national environment, with an attempt to spell out the tail risks of October Surprises in advance.
FRONT is a forecasting model for the 2026 midterms, run every Saturday from October 10 to October 31. It calls each of the 435 House races and 35 Senate races, gives a probability for each, and keeps every earlier run on the page so the calls can be judged after the fact.
The name is the method. Fundamentals: each district’s partisan lean, incumbency, the incumbent’s record of running ahead of the presidential vote, candidate quality, and local economic stress from the America Under Pressure state panel. Ratings: Cook, Sabato, Inside Elections, Decision Desk HQ and Split Ticket, converted to margins. Observed polls: district and state polls, with partisan internals discounted. National environment: the generic ballot, blended with a presidential-approval model fitted to past midterms and the special-election record since January 2025. Tails: fat-tailed errors that move together nationally, regionally and by state, plus a pre-registered register of how an October surprise would be handled.
Prediction markets are shown for comparison but carry no weight in the model.
Parameters come from the historical record, not from this cycle: the generic ballot’s final-week error in midterms since 2006, Senate and House polling error by cycle, the approval-to-seats relationship since 1994, and the calibration of the expert raters. The full memos, the inputs and the model code are kept with the project.
To Democrats (23 House, 5 Senate): North Carolina Senate (97%), Ohio Senate (special) (68%), Maine Senate (67%), Texas Senate (57%), Alaska Senate (56%), UT-1 (100%), CA-41 (100%), CA-3 (99%), CA-1 (99%), NE-2 (95%), CA-48 (91%), AZ-1 (83%), PA-7 (78%), CO-8 (76%), VA-2 (74%), NJ-7 (74%), PA-10 (73%), IA-1 (72%), AZ-6 (70%), MI-7 (69%), PA-8 (69%), IA-3 (68%), WI-3 (62%), NY-17 (59%), NC-11 (57%), MI-10 (56%), CA-22 (56%), IA-2 (52%).
To Republicans (9 House, 0 Senate): LA-6 (99%), TN-9 (97%), TX-32 (95%), TX-9 (94%), FL-9 (79%), AL-2 (73%), TX-35 (62%), ME-2 (57%), FL-22 (56%).
CA-22 (56% D), FL-22 (44% D), IA-2 (52% D), ME-2 (43% D), MI-10 (56% D), MI-4 (40% D), MT-1 (35% D), NC-11 (57% D), NY-17 (59% D), OH-7 (45% D), PA-1 (39% D), TX-35 (38% D), VA-1 (42% D), WI-3 (62% D), Alaska Senate (56% D), Iowa Senate (45% D), Texas Senate (57% D).