Hypersight is a multi-signal House and Senate model. Prior forecasts leaned too hard on polls alone. We blend public surveys, campaign finance, ads, and structural factors into a calibrated win probability for every district and Senate seat — then refresh as new data arrives.
Poll-only models fail in predictable ways: thin samples, house effects, late breaks, and races where the average never moves even as money and ads do. Single-signal stories (cash rankings, generic ballots, one flashy poll) overfit the last narrative.
Hypersight was built to be a rigorous multi-signal engine — accuracy- and recency-weighted polls where they exist, FEC-backed finance and ad pressure where they matter, and partisan structure as a baseline — trained and tested walk-forward so we never peek at the cycle we score.
ST-##
or Senate state for the cycle (e.g. 2026).
When all core inputs load, the blend is roughly:
Ratings come from Democratic win probability. They are confidence bands, not guarantees. Map colors use the same scale as these bands: blue for Dem favorites, brand green for toss-ups, red for GOP favorites.
We train only on prior cycles and score the next one — the same constraint a real forecaster faces. Headline House/Senate results:
On a 2024 holdout (model fit only through 2022), Hypersight correctly called about 95% of House and Senate seats in our feature set. Close actual races (two-party vote near 50%) remain the stress test — accuracy falls into the mid‑70s there, which is where multi-signal blends earn their keep.
Many public forecasts still start from a flat average of surveys. That “poll average” is useful — and it is only one channel. Hypersight accuracy- and recency-weights every poll, then blends finance, ads, and structure so a noisy mid-pack average does not get the final word.
Below are walk-forward holdouts (model trained only on prior cycles). In each case a simple two-party poll-average forecast pointed at the wrong winner; Hypersight’s multi-signal probability pointed at the eventual winner. Retrospectives on public features — not election-night posts.
Aggregated two-party poll share sat essentially at a toss-up / thin Dem edge (~50%). Hypersight’s blend put Democrats near ~31% to win. Mike Braun defeated the Democratic incumbent — the multi-signal call, not the poll average, had the winner.
Simple poll averages hovered just under 50% Dem (a soft Republican lean). The full stack pushed a clear Democratic probability. Democrats took the district — again matching Hypersight over the poll-mean favorite.
Survey averages put Democrats a hair over 50%. Hypersight stayed Republican on the blend of polls, money, ads, and lean. The seat stayed with the GOP — wrong favorite if you stopped at the poll average.
Two-party poll/forecast share averaged Democratic. Multi-signal probability favored Republicans. The Republican carried the district — a clean miss for a poll-mean call and a correct multi-signal one.
Available two-party forecast shares sat near 50%. Hypersight’s full blend pushed a clear Democratic probability and matched the razor-thin certified result.
In both cycles Nevada’s poll averages sat a tick under 50% Dem. Hypersight stayed firmly Democratic and the Democratic nominees won — another pattern where the mean of surveys understated the eventual winner.
No model calls every toss-up. Special elections, scandal, nominee quality, and late ad blitzes can move races after a snapshot freezes. Jungle primaries and multi-candidate fields need care. Photos and names come from FEC and public legislator data and can lag filings.
Hypersight is decision support — not a betting tip sheet. Use the live map for the current cycle, click a race for signal weights, and return here for methodology.