Forecast methodology
Akashic’s forecast does neither what the raters do nor what the quants do. We commit to one expectation, set by editorial judgment, and express it probabilistically — calibrated, honest about uncertainty, and scored in public on the same page as the result.
Every 2026 Senate, House, and Governor race is rated on an eight-point scale — tilt · lean · likely · safe for each party. There is no tossup: every race is rated with a favored side, with tilt the most competitive tier. The rating is a human call, grounded in the structural record Akashic already owns: 148 years of same-office results, the partisan lean of every geography, demographics, incumbency, and seat exposure. We do not chase polls, and we do not re-rate week to week. The conviction is the product.
Each rating tier maps to a baseline probability that the favored party wins, calibrated from the historical hit-rate of that tier across past cycles (Senate, House, and Governor races retro-classified by their prior-cycle margin). Where a predicted two-party margin is supplied, the probability is derived instead from the empirical forecast-error distribution, P = Φ(margin / σ), with σ the per-office spread below.
| Rating tier | Reading | P(favorite wins) |
|---|---|---|
| Safe | Not competitive | 95.9% |
| Likely | Clear favorite | 86.4% |
| Lean | Favored, not safe | 81.1% |
| Tilt | The most competitive tier | 58.0% |
Per-office forecast-error σ (the spread of actual minus predicted margin): U.S. House ±14.1 pts, U.S. Senate ±21.8 pts, Governor ±15.1 pts. These are large because a prior-cycle margin is a deliberately weak proxy for a real rating — a conservative upper bound on a well-rated forecast’s error. Calibration version backtest-1.
The chamber outcomes come from a 50,000-run Monte-Carlo over every contest. Each run draws a single national environment swing — s ~ N(0, ±7.8 pts) — applied to every race at once, plus an independent per-race shock. Modeling races as independent would manufacture an absurdly narrow, dishonest seat distribution; the shared national swing is what makes the spread honest. The variances are set so each race’s own win probability is preserved exactly, no matter how strong the national correlation. From the simulation come the control probabilities, the median seat count, the 80% and 95% ranges, and the tipping-point race.
The forecast is drafted privately, then locked and published once — ratings, margins, the simulated distribution, and this methodology, all frozen with a timestamp. A genuine error correction is a new version that preserves the original; the record must always show what we actually said. A forecast is never shown as a result: it carries a persistent prediction banner and a distinct, dashed visual treatment, and it never shares a surface with a live call.
When the results land on the same race pages, every locked call is scored automatically and kept permanently: the Brier score on the win probabilities (overall and by office), a calibration curve (did our 70%-favorites win about 70% of the time?), the directional record with the misses named, and the seat-forecast error — was the actual composition inside our 80% range? Each scored cycle is evidence for the next forecast’s credibility.
The forecast data is openly licensed and machine-readable, like the rest of Akashic. · Back to the forecast →