The Scenario Machine

The wave. It moves every race at once. The forecast was run 100,000 times at each of 31 steps, from a national House vote of R+6.6 to D+23.4; every race and chance at a step is that run's own, and nothing in between is estimated.

Published baseline. CTMM 2.2 — expert-free forecasting, smooth Senate calibration and fitted uncertainty, as of 2026-10-27. Arithmetic on the model's published numbers, not a re-run of the simulation. The chances beside the wave are the forecast's own, run whole at each national swing; once you set a Senate or House race, that chamber's are counted from the simulated nights that match it (below). No win probability moves with the group sliders or the starting points.

Hand edits. A result you set (a winner and the margin it wins by) or a Toss-up changes that race and counts it in the seat totals. A rating is a result at a representative margin, the middle of that word on our margin scale: Tilt 1.5, Lean 5, Likely 10 and Solid 20 points. A same-party general, a multiparty field the scenario does not restate, and a race whose call the forecast withholds cannot be edited.

Starting points. “Toss-ups for you” leaves uncalled the races the forecast rates a Toss-up. “2024 again” gives each race to the side its state, or its district on the 2026 lines, voted for in the 2024 presidential race, by that margin. “Markets” rates each race with a Polymarket price (as of 2026-10-27) on the site’s chance scale and leaves the rest as published. A wave is a position of the national control, at the precomputed world nearest it. A race a starting point sets is a premise, not a result: it does not move the others or condition the counted chances; only the races you set yourself do.

Results move the others. In the chamber you set a result in, every other race moves by its own count: the forecast publishes the simulated margins of 20,000 of its Senate nights and 10,000 of its House nights, and the page keeps the nights in which your race came out within 1.5 points of your margin (wider, and said so, when fewer than 1,000 come that close), moves each the rest of the way by how the other races move with yours in those same nights, and averages. A governor race, and a chamber you set nothing in, move by the national shift your results imply instead: part of every race's error is shared nationally, and each simulations file states how much (a race's own spread and the spread all races share; governors' is measured the same way from the model's September 19 draws). A result implies a shift of (your margin − the race's margin before it) × that shared share, combined across results as a precision-weighted average with the published forecast as one more observation, and applied at the national control's nearest precomputed world. Every race you set stays where you put it. A what-if, not a forecast: no published chance changes.

Win chances. The forecast publishes who wins each Senate race in every one of its 100,000 simulations, and each House district either side wins at least 5,000 times in a fixed sample of them (edition of 2026-10-27). When you set winners there, the page keeps the simulations in which they happen and counts again; when you set a margin, it keeps the nights in which the race came out close to it, as above. That is a count of the published simulations, not a new forecast, and it is shown however few match: under 1,000 each figure is given as its count, and under 100 the page marks every figure and shade as rough. An edition published before October 26, 2026 holds winners only, and there a margin counts as its winner. Governor races have no published simulations, so they are not counted.

Sources. Race and education, age, urbanicity and gender use sourced crosstabs. Party identification, approval and ideology use modeled priors, not surveys of each race.

Shares move together. Within each dimension the selected turnout shares are renormalized to exactly 100.0%, so raising one group lowers the rest. Dimensions combine additively and can describe the same voters twice: this is transparent scenario arithmetic, not a synthetic joint-voter microdata model.

Race and education. White voters split by degree; Black and Hispanic/other voters remain whole because no complete race-by-education table is published.

Not admitted. Household income, complete education, registration, union membership, homeownership, marital status, religion and region/media market — each 0/506 qualifying race-level scenario rows.

How the scenarios are built →