We never rely on a single forecaster. Every race rating starts as an average of averages across Cook Political Report, Sabato's Crystal Ball, Inside Elections, Decision Desk HQ, RealClearPolitics, Silver Bulletin, and prediction markets (Kalshi). This keeps our projections non-derivative of any one source and smooths out individual house effects.

Sources are weighted by their historical accuracy, and we then adjust for systematic polling error — the documented tendency of state polling to miss in one direction. Examples baked into our current ratings: Maine polling missed Susan Collins by ~13 points in 2020; Ohio polls have understated Republicans by ~4 since 2016; New Hampshire polling routinely overstates Democrats late.

That's why our projected margins sometimes differ from the raw polling average — the raw average has been wrong in predictable ways.

Chamber-control odds come from 20,000 simulated elections. Each simulation draws a shared national environment error (±3 points, correlated across all races — because polls tend to miss in the same direction everywhere) plus an independent race-level error (±4–5 points). A race's rating implies its expected margin; the simulation counts how often each party reaches a majority.

The simulation is seeded, so the odds are stable — they only move when we change a rating.

State detail maps apply our statewide forecast to certified 2024 county returns as a uniform swing: if we project a state 4 points left of its 2024 result, every county shifts 4 points left. It's a simplification — real swings vary by county — but it's transparent, reproducible, and grounded in real returns rather than modeled turnout.

Ratings move only when the evidence moves: a candidate entry or exit, sustained polling shifts, fundraising reports, or an aggregator consensus change. Every move is logged with its reasoning in the Latest Ratings Changes feed and explained in each race's detail panel.

Rating scale: Safe (12+), Likely (6–11), Lean (2–5), Tilt / Toss-up (<2) — based on projected margin.