About the numbers
Trust, but verify.
How the forecast earns it: backtested before it touches 2026, archived every month after, and every number traced to a source.
Proof, not promises
Calibrated on the past. Frozen before November.
A forecast is only as good as its calibration. Before the model touches 2026, it must survive 2018 and 2022: how close our probabilities came to the actual results (Brier score and log-loss), reliability curves showing that races we called 70% actually landed ~70% of the time, and checks that our 90% ranges — the ranges we'd bet on — captured the real result about 90% of the time. The full validation report — every number, every miss — ships with the model.
- Backtested on 2018 and 2022 before forecasting 2026
- How close our probabilities came to the results — published
- Blend schedule frozen and pre-registered — no October re-tuning
- Per-race calibration is the evidence; one chamber outcome is an anecdote
Wayback · forecast history
Every forecast, archived. Every revision, visible.
Drag the scrubber — tracking begins when the model launches. Monthly snapshots, locked at publication, each revision dated and explained.
Snapshot
Sep 2026
No snapshot recorded — by design. Backtests on 2018 and 2022 will seed the archive at launch, then every live forecast is written to an immutable log. Nothing appears here until it is real.
Cited sources
Every number traces to one of these
FEC API
Candidate filings, fundraising totals, and committee financials — the raw material for candidate-quality scoring.
Public domain (U.S. federal data)
VoteHub
Structured election results and race metadata, kept current through Election Day.
CC BY 4.0
FiveThirtyEight polling archive
20,466 historical polls (2008–2024). Vendored in full — this is the dataset our correlated-error structure is fitted on, not asserted from.
CC BY 4.0
MIT Election Data & Science Lab
County- and precinct-level returns used to backtest the model on 2018 and 2022.
CC0 (public domain)
U.S. Census — American Community Survey
District and county demographics (5-year estimates) feeding the fundamentals layer.
Public domain (U.S. federal data)
Cook Political Report — Partisan Voting Index
District/state partisanship baseline. Used with attribution; race ratings shown for context are Cook's own judgment, not ours.
Licensed product (not open data)
District boundaries shown are the 119th Congress (2024 elections). Mid-decade redraws and litigation across 2025 — in Texas, California, North Carolina, Florida, Tennessee, Utah, and Louisiana — may change 2026 lines; the production map will ship the certified 2026 geometry.