Florida

What we set out to do

Two weeks before Election Day, our Oct 20–23 Florida poll showed Republicans ahead. At the same time, Florida's own early-vote reports showed Democrats and voters with no party affiliation turning out at rates far above 2022. Polls model the likely electorate on past turnout. If this year's voters really are different, that model is wrong.

So we went back into the field to answer one question: is the electorate changing, and if it is, what does that do to the race? We were weighing a model change before the new surveys came in, and fielded them to see whether the evidence called for one.

When we changed what. Both new weighting targets were set after fieldwork and after we had seen earlier results: the party target on Oct 31 and the 2024-vote target on Nov 1. We had written down a rule in advance for changing the turnout target, and we dropped it in favour of the projection below. Our first version of that projection also counted about one day of ballots twice; we caught it and corrected it before publishing. Judge the change knowing all of that.

Three surveys

1. Oct 20–23: a statewide poll plus five congressional-district polls. 4,823 likely voters, drawn from the voter file and interviewed by text, in English.
2. A re-interview panel: 540 completes, Oct 30–Nov 1, from Oct 20–23 respondents who hadn't voted yet when we first reached them. 534 link to their first interview. The panel leans toward the most likely voters, so we use it to compare people with themselves, not to estimate the race.
3. Oct 30–31: a fresh statewide poll plus five district fills. 2,898 likely voters, interviewed by text in English and Spanish (278 answered in Spanish). This is our headline poll.

Our headline numbers come from the Oct 30–31 poll alone. An experimental estimate that combines the two statewide polls is on the data page, labelled as experimental.

What the early vote showed

These are official counts from the Florida Division of Elections and the state voter file, not survey answers.

• Ballots by registered party. At the same point before Election Day as 2022, Democratic ballots were up 14% and NPA ballots up 14%, against 2% for Republicans. That's despite fewer registered Democrats and NPAs than in 2022. Per registered voter, Democratic turnout so far is up 38% and NPA up 36%; Republican is down 6%.
• Returning voters. 451,000 of this year's early voters were registered for the 2022 midterm and skipped it, more than twice the comparable 2022 figure. That ratio is somewhat overstated: people who voted in 2022 and have since left the rolls aren't in today's file. Among early in-person voters through Oct 30 who skipped 2022, 86% voted in 2024. These are mostly presidential-year voters showing up in a midterm.
• Who they are. Returning voters are disproportionately Black, Hispanic and young.
• Other states. Voters returning after skipping the last midterm are up 1.6–2.1× in our voter files for North Carolina, Georgia and Clark County, Nevada. In North Carolina and Clark County, where registration has party, they now run about 2-to-1 Democratic. The counts are partial and the methods differ by state.
• Ballots moved, not just added. Of Democrats who voted on Election Day in 2022, 28% have already voted early in person this year, against 12% over the whole early-voting period last cycle. Republicans are at 20%, their usual level. So Democrats have spent more of their Election Day vote, and more of Republicans' is still to come. Two cautions: the baseline comes from one pair of elections (2018 to 2022) and covers the whole early-voting period, while 2026 is counted through Oct 30, so both parties' figures will rise before Election Day.

Update 1: who we expect to vote

Our earlier model expected the electorate to be 21 points more Republican than Democratic by registered party. We replaced that with a projection built from this year's early-vote counts, completed with 2022's late vote, done two ways and averaged.

The result: 46% Republican, 33% Democratic, 19% NPA, 3% minor parties: R+13.

Against registration, each party's share of voters minus its share of registered voters:

Against the last midterm, Republicans out-vote their registration by far less, and Democrats out-vote theirs for the first time in these three elections.

• How well this works. We could test these projections on only one past pair of midterms (2018 to 2022). Every version was too Democratic, by 4–6 points; ours missed by 5. The evidence favours our projection over a purely additive surge, and the bigger risk is an electorate more Republican than ours.
• The range. If the surge is all new voters, the electorate is R+9 and both races are within a point (Senate about even, Jolly +1). If it's mostly timing, it's R+16 (Moody +6, Donalds +4). Between those, Nixon pulls even at about R+8 to R+9 (by interpolation, not a separate run).

Update 2: how we account for 2024 votes

Most pollsters now adjust their samples so respondents' remembered 2024 vote matches the official result. Our respondents remember voting Harris about 6 points more than Florida actually did. Left alone, that would tilt the poll toward Democrats; the adjustment pulls it back.

That adjustment matters a lot here: turning it off moves the Senate race about 8 points. So we tested it every way the records allow.

1. Does this year's electorate have a different 2024 history? We combined county and precinct 2024 results with the voter file's record of who voted in 2024 and this year. By place and party, this year's projected voters with a 2024 vote went about 42.7% Harris (43.3% counting all projected voters), against 43.4% statewide. The change is mostly in which parties are turning out, which Update 1 already captures.
2. Are our respondents from bluer neighborhoods than they should be? We found no sign of it. Within their county and party, respondents live in the same kinds of precincts as the people we texted who didn't answer, and as this year's early voters. Limits: five counties couldn't be tested, precincts are current addresses, and mail voters were left out.
3. Are the Republicans who say they voted Harris real? Many are, but not all. In the records, 35% of them voted by mail in 2024, against 22% of Republicans who say they voted Trump and 43% of Democrats who say they voted Harris: about 60% of the way toward the Democrats. Up to about 40% of them could be misreports; the national misremembering rate below implies about 28%. Primary records neither confirm nor refute their answers.
4. Do people misremember their 2024 vote? Yes, measurably. A published study (G. Elliott Morris / Verasight, Oct 22, 2026) asked the same 2,028 people twice: 3.2% of Trump voters (22 people) now say they voted Harris, against 1.4% of Harris voters (11 people) the other way. The counts are small, so the rate is uncertain, and it's national, not Florida.

So we set the 2024 target to the records-based estimate for this year's projected voters (42.7% Harris), then adjusted it for that measured misremembering, giving 44.0%. Both inputs come from outside our poll. Records can't see how individuals voted, only places and parties.

Independent review

AI models from several companies reviewed the data and methods: Claude Opus and Claude Fable (Anthropic), Codex (OpenAI) and Muse (Meta). Each could read every output and the code but change nothing. Their reviews are advice, not proof.

• Blind round. All polling-error history was removed, so no reviewer could lean on the past decade's misses in either direction.
• Complete-evidence round. Four fresh sessions read all the evidence at once and answered one question: will the remaining vote change the party makeup of the electorate?

In the complete-evidence round:

• all four were 75–95% confident the remaining vote will be more Republican than 2022's, mainly because Democrats pulled Election Day votes forward;
• their final electorates ranged from R+11 to R+14, centred near ours;
• all four accepted the misremembering adjustment;
• three would shade the result up to 1 point toward other Florida polls; we didn't, because our method doesn't use other polls;
• their own Senate finals ranged from Moody +3 to Moody +5.

The result

Leads are computed before rounding, so they can differ by 1 point from the shares shown. Without a 2024-vote target the poll reflects our respondents' remembered vote, which runs well to the left of Florida's result. The records suggest much of that is who answers polls and how people remember, but they can't fully separate the two.

Margin of sampling error: ±2.1 points per candidate, ±4.1 on the gap. That covers sampling only, not uncertainty about who turns out or about our model. Both races are within the margin.

Limits

• Text-to-web polls reach people who answer texts.
• Registration isn't a vote.
• The turnout projection could miss, most likely in the Republican direction.
• The misremembering rate comes from a small national panel, not Florida.
• Records can't see how individuals voted, only places and parties.
• Our weighted sample is about 79% people who say they've already voted, against about 59% of the projected electorate in the records; a check weighting to both party and voted status moved the Senate lead by a tenth of a point.