What PoliAgg is
PoliAgg collects polls and prediction-market data for US elections from public sources and presents them in one place — polling averages, model forecasts, pollster scorecards, and prediction-market odds, refreshed daily. A daily report and a Sunday digest analyze what moved and why; every report is machine-drafted from the data on file, then checked against that data before it publishes. We endorse no candidates and take no advertiser input on coverage.
- Wave collapse — when one survey produces multiple matchups (e.g. "X vs A" and "X vs B"), they are collapsed into one wave so the survey doesn't get N× weight.
- Recency decay — exponential, 14-day half-life. Polls a half-life old count half as much as today's.
- Sample weighting — √(n / 600). Polls without a reported sample size fall back to inferring n from the reported margin of error.
- Pollster weighting — inverse of mean error from final-stretch polls in past cycles, clipped to [0.25, 4.0]. Default 1.0 for pollsters without enough historical data.
Mean error and bias are computed on each pollster's final-21-day polls in past elections vs the actual two-way margin. Lean is derived from the signed bias (≥2pt and n≥5).
- Cross-cycle pollster weighting helps marginally (0–0.3pt MAE) — most error is cycle-specific bias that didn't exist in past data. The 2024 cycle had ~3pt industry-wide D-bias on final polls.
- Multiple-matchup polls store all variants but the headline average reduces a candidate to one number per survey (mean across matchups). This dampens distinct hypotheticals.
- Weights are deterministic (inverse mean error). A Bayesian state-space model would handle uncertainty better; that's planned for a later phase.
The site aggregates from many sources, each with its own license and attribution requirements. The canonical record is /sources; key contributors:
- Wikipedia — race calendars, poll tables, ratings · CC BY-SA 4.0
- Polymarket / Kalshi / PredictIt / Manifold prediction markets
- GDELT Project · CC BY 4.0
- OpenSecrets · CRP donor data
- Federal Election Commission Open API
- Ballotpedia · CC BY-SA 4.0
- State Secretaries of State (28 states)
- AAPOR Transparency Initiative member list
- External race forecasters · 10 raters via Wikipedia Predictions tables
Coverage spans 2026 US Senate, Governor, and House races, plus 2028 Presidential ratings, ballot measures, and topic trackers — alongside historical 2020/2022/2024 cycles for pollster scorecards. Depth varies by race: where polling is thin (many House and Presidential contests), ratings lean on PVI and campaign fundamentals rather than a deep poll average.