How to use the 2026 Election Atlas

Following an election means keeping track of numbers that answer different questions. Polls ask people how they plan to vote. Early-voting reports count ballots. Past results show how a state has voted before.

The American Optimists Election Atlas puts those numbers in one place for 40 Senate and governor races in 24 states. Pick a state to compare its polls, election history and early-voting pace, or see how our model changes when you try different assumptions.

Start with a state

Open Projections and choose a state. You'll see an estimated Democratic–Republican margin, a range of possible outcomes and, where polls are available, a polling-only comparison. Select a race to see the historical results and demographic data behind the estimate.

“D +3” means a three-point Democratic lead within the combined Democratic and Republican vote. Other candidates aren't included. The conditional lead percentage tells you how often Democrats lead in that two-party comparison; it isn't the candidate's chance of winning the election. Independent candidates, ranked-choice voting and runoffs can change who wins.

Try the ten-, twenty- and forty-year training windows to see how much the choice of history matters. Under Specification, switch between political history and the model that also uses demographics.

The scenario controls let you change assumptions about education, population, age, income, unemployment and the national political environment. Your changes affect the comparison on your screen, not the published forecast. They show how the model responds to an assumption; they don't prove that a demographic change causes a voting shift.

Find a county or precinct

Open County & precinct, search for a county and choose the matching state. Then select Precinct results to search the available precinct labels or numbers. Use the year and office filters to narrow the results, and download the displayed records as a CSV. On a phone, the Views button opens the navigation menu; wide tables scroll sideways.

Coverage varies by state and year. Some reporting units combine absentee or early votes, and precinct boundaries can change between elections. Missing records mean we lack verified data, not that nobody voted. Precinct results are historical evidence, not precinct-level forecasts.

Compare early voting day by day

Open Early voting, choose a state and select the years you want to compare. The chart lines them up by days until Election Day: fourteen days out against fourteen days out, then thirteen against thirteen, and so on.

You can view cumulative totals, reported daily counts or changes between consecutive days with observations. Compare the same voting method and ballot stage. A requested mail ballot, an accepted ballot and an in-person vote are different things. A missing report leaves a gap in the chart, not a zero.

Reporting delays and changes in voting rules can affect the shape of the lines. Presidential and midterm turnout also differ. And a higher early-vote count doesn't tell you which candidate is ahead: even a voter's party registration doesn't tell us how they voted.

What the Nevada comparison shows

The Nevada post shared by Nick Field is a useful example. In late-evening reports eight days before Election Day, Ralston's 2026 tracker put registered Democrats ahead by 24,395 ballots. His 2024 tracker had Republicans ahead by about 40,000.

That's a roughly 64,400-ballot change in the party-registration gap. It's encouraging turnout evidence for Democrats, but it isn't a swing in counted votes for candidates. Nonpartisan voters, crossover voting, ballots still to come and mail-reporting cutoffs all matter.

These analyst-reported totals give us context. They haven't been added to the Atlas's official-source charts or used to adjust the forecast, and they can't settle Nevada's governor race or establish a national trend on their own.

Where the data comes from

Historical results come from the MIT Election Data and Science Lab, the county archive assembled by Carlos Algara and Sharif Amlani, and official state election results. The archive covers 1986–2025, with some earlier reference elections. It doesn't include every election or every county.

Demographic data comes from the U.S. Census Bureau, including the decennial census and American Community Survey. We also use population and income data from the Bureau of Economic Analysis and unemployment data from the Bureau of Labor Statistics.

Early-voting reports come from state election authorities. Each poll links to its original release and the available details about how it was conducted. Campaign-finance data comes from the Federal Election Commission and available state disclosure systems.

In Sources & coverage, you can follow those links, check retrieval dates and see what's missing. The research and market-comparison views offer other ways to examine a race. Agreement with a betting market doesn't establish that our model is right.

How often it updates

A refresh is scheduled for 9 a.m. Eastern each day. It checks for new polls and supported early-voting reports, reruns the models and publishes after validation. If the run fails, the last working version stays online.

Check both the dashboard's snapshot date and the date on the source you're using. States report on different schedules, and coverage varies by race. A new dashboard snapshot doesn't mean every source has new data. The scheduled run currently depends on the publisher's Mac and Codex being available.

Polls and early-voting counts update their own comparison views. They aren't direct inputs to the model built from election history and demographics.

How much confidence to place in the model

The model trains on 1,171 eligible Senate and governor contests across all 50 states. To test it, we use earlier elections to predict later ones, holding out entire election years.

For the forty-year version, the political-history model's average margin error was about 14.05 percentage points. Adding demographics produced an error of 14.57 points; a simple baseline using the prior presidential election had an error of 16.59 points. Adding demographics didn't improve that test, so political history is the default. You can inspect both.

Those are large errors. The tests also use historical data that may have been revised since the election, so they don't recreate exactly what an analyst knew at the time. Treat the Atlas as an experimental tool for comparing evidence, with substantial uncertainty around its estimates.

Open the Election Atlas and start with a race you follow. Compare its polls with the historical model, check the source dates, and try a different training window to see how the estimate changes.