Skip to content

Methods

Reading the map

Each state or district is coloured by its polling average: every pollster's latest poll with fieldwork in the last 21 days, one poll per pollster. For past elections the same is shown for the final 21 days before election day, beside the result.

  • Corrected or plain. The corrected average adjusts each pollster for how far it missed in past elections of the same kind (midterm or presidential year). It is the default on the 2026 map, and the plain average is one click away. In our test on every election from 2006 to 2024 (experiment E003, in the lab), the correction did not beat the plain average.
  • Margin of error. With it on, purple marks a race whose lead is smaller than the average's 95% sampling margin of error. That covers sampling error only: real polling misses have averaged 4–6 points. With it off, every race goes to whoever leads.
  • House districts without polls are drawn faded and show a prediction: the district's 2024 result shifted by the national swing (the generic ballot, D+7.0, against 2024's national House vote of R+2.9). The method passed its test on four past midterms (experiment E006, in the lab). How often a predicted winner actually won, by predicted margin: under 5 points, 52%; 5 to 10, 64%; 10 to 20, 88%; 20 or more, 99.5%. So a prediction under 10 points is close to a coin flip, and is counted as close. No prediction is made in AL, CA, FL, LA, NC, OH, TN, TX, UT, which have new district lines for 2026, so their 2024 results do not carry over.
  • Balance of power counts who leads where, not a forecast: a seat in a close race goes to neither side.

Live results and projections

On election night the map shows official counts, read from each state's own results site every minute or so. Every file read is kept, so each number can be traced to what the state published. States whose sites cannot be read automatically, or that publish no statewide count on the night, fill in from certified results later.

  • A lighter shade means a candidate is ahead in the count so far. Solid colour is a projection from our own model, which is experimental: it was replayed on the 2024 count in 20 states, in many reporting orders, without a wrong call, but it is not a network or AP call. House races are not projected.
  • Before 6 p.m. Eastern on November 3 the states' feeds hold their own test data, so live results are not shown.

How a poll is scored

For a race between one Democrat and one Republican:

poll margin   = poll D% - poll R%
actual margin = actual D% - actual R%
signed miss   = poll margin - actual margin   (positive = overstated the Democrat)
absolute miss = |signed miss|

Margins are computed from each candidate's share, never copied from a source's own margin column. Races that are not one Democrat against one Republican are not scored this way; they are counted and reported as skipped.

Which polls count

  • Fieldwork ended within 21 days of the election.
  • One question per poll per race (the largest sample), so a poll asking several versions counts once.
  • Near-duplicate releases (same pollster, date, race and numbers under two IDs) are flagged, never counted twice.

Uncertainty

Ranges are 95% intervals from a bootstrap that resamples whole races, because polls of the same race share that race's outcome. Pollsters with fewer than 10 scored polls are labelled as small samples.

Point in time

Historical replays only use polls that archived evidence shows were public by the date in question, with the numbers that evidence showed. Replays that use later listings or fieldwork dates are reported alongside, to show how much hindsight would flatter them.

Sources

  • ABC News / 538: pollster-ratings raw_polls.csv (CC BY 4.0)
  • ABC News / 538: poll-page CSVs, archived 2025-03-06 (license unconfirmed; research use)
  • ABC News / 538: election-results (no license stated)

Scored polls: 9,457 · pollsters: 495 · generated 2026-10-24T17:50:06+00:00