Methodology

How the ratings, poll averages, seat counts and what-if scenarios are calculated. Method version 2. Data last refreshed Oct 25, 2026, 1:31 PM UTC.

Data and freshness

• Ratings, polls and nominees come from Wikipedia’s 2026 election pages under CC BY-SA, which compile ratings from Cook Political Report, Decision Desk HQ, FPO, Fox News, Inside Elections, NBC, RealClearPolitics, Sabato's Crystal Ball, Silver Bulletin, Split Ticket, The Economist and list individual polls with links to their original releases. We don’t scrape the NYT, Cook Political Report or RealClearPolitics; we link to them.
• Candidate finance for House and Senate comes from the FEC’s OpenFEC API.
• Refresh Every 6 hours. The daily update runs at 6am Pacific.
• We only see changes after Wikipedia’s editors do, so a new rating or poll can lag its source by hours.

Race ratings

A race’s rating is the average of its raters’ ratings, not a poll-based model. House and Senate races have 10 raters and Governor races have 8.

• Scoring. Safe or Solid = 3, Likely = 2, Lean = 1, Toss-up = 0. Democratic ratings are negative and Republican ones positive. “Tilt” ratings count as Lean.
• Average. All raters count equally, including the model-based ones alongside the expert raters. One rater moving one step changes the average by 1/10, which is 0.1 for a 10-rater race, so several raters must move before a race changes category.
• Rounding. The average rounds to the nearest rating. A tie exactly halfway between two ratings goes to the less decisive one, the same way for both parties, so an average of ±0.5 is a Toss-up, and ±1.5 is a Lean and not a Likely. So a race is a Toss-up when its average is between −0.5 and +0.5.
• Competitive races are those rated Toss-up or Lean. Pennsylvania and Maryland governor are also always listed as competitive by choice, even though raters call them Likely or Safe.
• Unrated seats that are not in the raters’ tables are counted as Safe for the party that holds them. Seats currently held come from each chamber’s current composition on Wikipedia, which puts the House at 214 Dem today, so the House comparison is approximate because of vacancies and new and redrawn districts.

Independent candidates

Raters only rate Democrat against Republican, so we look at the polls when an independent is in the race. An independent appears on a race when polling within 8 points of the leader or leading, and the race counts as competitive within 5 points. If the independent leads the poll average by 2 or more points, the race is Lean or Likely Independent, with Likely at 6 or more, and a gap under 2 points either way is a Toss-up. Otherwise the raters’ rating stands.

Editorial overrides

We can replace the raters’ average for a specific race. An override applies everywhere on the site, the race page says so, and the daily update reports the move as “in our own rating.”

• senate mi Toss-up. Polling is too close to call, so we rate this a Toss-up.

Seat forecast

• Counts. Each race goes to the party its rating leans toward; Toss-ups are counted separately, not split. Independent-favored races are counted separately too.
• Whole chamber. The totals include every seat, meaning the individually rated races plus the unrated safe seats.
• Net shift. “Dems +X to +Y seats vs. the N they hold now” compares the Dem-leaning count with the seats Democrats hold today. The low end gives every Toss-up to Republicans and the high end gives them all to Democrats.
• Majority math. Distance from the majority, which is 218 in the House and 51 in the Senate, uses the Dem-leaning and Rep-leaning counts with Toss-ups left out.

Poll averages

Poll averages are shown next to the ratings but do not change the ratings, except in the independent-candidate rule above. Each general-election poll gets a weight, and the average is the weighted mean per candidate.

• Last 14 days. When a race has at least 5 polls from the last 14 days, only those count and older polls are left out. With fewer, older polls stay in and are down-weighted by the rule below. The window counts back from the last data refresh.
• Recency. A poll’s weight halves for every 14 days it is older than the newest poll in the average.
• Sample size. Weight grows with the square root of the sample and is assumed to be 500 when not reported.
• Population. Likely voters 1, registered voters 0.9, all adults 0.8, and 0.8 when not stated.
• Pollster quality. Tier A counts 1, tier B 0.8, anything else, which is tier C, 0.55. A poll from a pollster or sponsor tagged Republican or Democratic counts at 0.5 of that.
• “High-quality polls only” on a race page keeps only nonpartisan pollsters in tiers A and B, and the 14-day rule is applied to what remains.

The tiers are our own editorial defaults and not an official rating. Tier A is new york times, siena, marist, quinnipiac, emerson, marquette, ap-norc, ap votecast, cnn, ssrs, fox news, beacon research, shaw & company, washington post, abc news, nbc news, cbs news, wall street journal, selzer, monmouth, pew research, suffolk, university of new hampshire, unh, mason-dixon. Tier B is yougov, ipsos, morning consult, surveyusa, university, college, public policy polling, echelon, data for progress, civiqs, swayable, atlas intel, atlasintel, tipp, and any university or college poll also counts as B.

What-if scenarios for a polling error like 2024

The bar under the page header switches the whole site to a scenario in which 2026 polls miss the way 2024 polls did, in the same direction. The choice sticks as you move between pages. Neither scenario is a prediction; they show how sensitive the forecast is to a systematic polling miss.

• 2024 polling error, national. Every race moves 2.7 points toward Republicans. 2.7 is the measured amount 2024 polls overstated Democrats, with about 2.7 points in competitive states per ABC/538, and 2.7 points across all offices in the final two weeks per the AAPOR task force.
• 2024 polling error, by state. Each race moves by its own state’s 2024 error in the table below. A state with no reliable estimate uses the national figure, and an estimate built from fewer than 5 polls is averaged with the national figure.

How a shift is applied

• Poll averages. Half of the shift goes to each side, so the Republican gains and the Republican’s opponent loses. Individual polls and the chart are never altered.
• Ratings. Every 3 points of polling error moves a race one rating step toward Republicans, so 2.7 points is 0.9 of a step. The shift is added to the raters’ unrounded average, then rounded as above. This applies to every race, including ones with no polls, because the error is assumed to be systematic. The 3-points-per-step figure is our assumption and is roughly the margin separating Toss-up, Lean and Likely.
• Left alone are independent-led races, meaning Lean or Likely Independent, and unrated safe seats.
• Races listed as competitive only by rule, for example Pennsylvania governor, stay listed in the scenarios. A race that becomes a Toss-up or Lean under a scenario joins the competitive list but may have no polling details, since those are only loaded for races competitive today.
• The daily update and the forecast without a scenario always reflect the current forecast.

State-level 2024 polling error

We computed each state’s error ourselves, because no complete published table was available. The error is the certified margin minus the final polled margin, both Harris minus Trump, from Wikipedia’s statewide polling and results tables. The polled margin is the 538 final average where Wikipedia lists one, or else RealClearPolitics or 270toWin, otherwise the simple average of polls in the last three weeks, only when there are at least 3. Negative means polls overstated Harris. Swing-state errors average about 2.3 points, in line with published figures. One poll is left out. Selzer & Co.’s final Iowa poll had Harris +3 in a state Trump won by 13, a miss of about 16 points that would by itself distort Iowa’s error. Both published Iowa averages include it, so Iowa uses its individual polls without it. Only 3 qualify, so the scenario averages that with the national figure.

23 of 50 states have an estimate. The rest use the national −2.7. These are historical and don’t change with each refresh.

Early and absentee voting

We show who has already voted. That is the share of returned early and absentee ballots cast by registered Democrats, Republicans and everyone else, next to 2024 and 2022 at the same point before Election Day. It does not change any rating, seat count, scenario or forecast, and it shows who has voted, not how they voted.

• Where it lives. The home page keeps a summary of the states we can compare at the same point. The Live Voting Results page has a map, every state and the day by day trend. Race pages link to their state there instead of repeating the charts.
• Only states that record party. Party comes from the voter’s registration. States without it, such as Georgia, Michigan, Ohio, Texas, Virginia, Wisconsin, show a ballot count only and are purple on the map. States with a party split today are Pennsylvania, North Carolina, Maine, Iowa, Nevada, Kansas, Nebraska, Colorado, Oklahoma, Maryland, Oregon, Arizona, Connecticut, Delaware, Idaho, New Jersey, New Mexico, Rhode Island, South Dakota, Utah, Wyoming, Florida.
• Same point before Election Day. This year’s numbers run through the latest complete day with returns. 2024 and 2022 are counted through the same number of days before their Election Days, so the comparison is like for like. A ballot counts on the date it was returned, and only accepted ballots count for North Carolina and Maine. Pennsylvania excludes cancelled ballots. States with no earlier year to match are shown separately, with this year only, and left out of the comparisons.
• Home page summary. It adds every state that can be matched at the same point and compares the total with 2024 and 2022 at the same number of days before Election Day in each state. States are at different points because each is counted through its latest complete day. The states that publish party are not a sample of the country.
• Live Voting Results map. It colors each state by which party is ahead among ballots returned, using the registered Democratic minus Republican gap. A deeper color is a bigger gap and a hatched tile means only part of the state. Purple states report a ballot count with no party split, using civicAPI for the count and the UF Election Lab for 2022 turnout and the 2024 early vote. Click a state to jump to its numbers. The headline adds every ballot we can count, with or without a party split.
• Trend chart. Click a state’s 2026 bar to see its ballots by party build up day by day against 2022 or 2024, counting down to Election Day. “So far” stops at today and “To Election Day” shows how the earlier year finished. A table underneath gives each party’s ballots and its share of all ballots now and at the same point in the earlier year. This year’s line comes from the state’s own daily record where it has one, and otherwise from civicAPI’s daily history, so it can start later or look steppier. A state with no earlier year daily record shows 2026 only.
• States with a dated record. Pennsylvania, North Carolina, Maine and Clark County list each returned ballot with its party and the date it came back, and Iowa’s archive keeps every daily report. We keep only daily totals by party and district, never voter names or addresses.
• civicAPI. For Arizona, Connecticut, Delaware, Idaho, New Jersey, New Mexico, Oklahoma, Oregon, Rhode Island, South Dakota, Utah and Wyoming, and as the current count for Kansas, Nebraska and Florida, we read civicAPI. It is a free public feed, free for commercial use with attribution, that compiles each state’s and county’s counts by party registration. A snapshot is taken once a day and counted through the day before. It keeps this year only, so these states have no same-point comparison with 2022 or 2024 yet. The finished 2024 total from the UF Election Lab is shown for reference where it exists. These states are left out of the same point totals and shown separately, with one combined bar for every state that reports party. Maine’s own file is sometimes posted late, so when civicAPI’s statewide count is newer and more than 1% higher, the difference is added on the day it covers and Maine’s district figures are then unavailable. If civicAPI is overloaded the run retries and then keeps the last good numbers. We also keep our own daily record of each state’s running total, which feeds the “since the last update” line.
• Maryland, mail ballots only and approximate. The State Board of Elections overwrites its daily mail-in sent and returned report, so 2022 and 2024 come from the days the Internet Archive saved a copy, about eight in each year and each stamped with an as of time of 6 PM, with the days between smoothed in a straight line. That makes Maryland’s same-point comparison approximate. This year’s count is the live report itself. In-person early voting is a separate file and is not included. Florida and Oklahoma were checked the same way and left out. The archive saved only a few Florida readings and none for 2022, and there is no Oklahoma mail-ballot history by party.
• Colorado. This year’s count is civicAPI’s, but 2022 and 2024 come from the Secretary of State’s own weekday ballot-returns releases, a spreadsheet of ballots returned, mail and in person, by party and county as of 11:59 PM the previous day. The archive of past releases starts about two weeks before Election Day, so Colorado is compared with 2022 and 2024 from 14 days out and shows this year only before that. There is no release on weekends, so a missing day is filled in by spreading the next release’s increase evenly across the gap.
• UF Election Lab. Iowa’s Secretary of State now blocks automated downloads, and Kansas and Nebraska hand their files to researchers, so for Iowa this year, and as a backup for those two, we read the free CSV downloads of the UF Election Lab early vote tracker run by Dr. Michael McDonald. It compiles each state’s file into ballots voted, mail accepted plus in person, by registered party, shared under a Creative Commons non-commercial license. UF stamps the day it compiled the file and we count it through the day before. For Iowa, 2024 and 2022 still come from the state’s daily archive, so the totals can differ slightly in definition. For Kansas and Nebraska UF has no daily history, so 2024 is shown as a finished total with no 2022 and they are left out of the combined charts.
• Nevada is Clark County mail ballots only. Nevada’s Secretary of State site blocks automated downloads, Washoe County publishes no daily files for earlier years, and Clark’s 2022 in-person early voting files are no longer posted. So Nevada shows Clark County’s mail ballots, about 70% of the state’s vote, by registered party. Only valid ballots count, so those returned undeliverable or missing a signature are left out. Clark logs ballots in batches, so day by day comparisons are rough.
• An indicator, not a forecast. Early voting by party is context only. It is not used in any rating, seat count, scenario or the daily forecast.
• Read with care. 2024 was a presidential year and 2022 a midterm, early voting rules and habits change, mail and in person voters differ, and the first days of voting are a small, unrepresentative sample. A chart warns when there are under 2,000 ballots. The numbers refresh with each data refresh.

Betting markets

• Sources. Prices come from Kalshi and Polymarket, read from their free public data feeds with each data refresh. We show who wins House and Senate control and each toss-up race. For governors we use Polymarket’s market on which party holds more governorships, because Kalshi has none. Races a platform doesn’t list show only the other platform.
• Reading a price. A contract that pays $1 if a party wins trades at that party’s chance, so 64¢ reads as 64%. The two sides of a market add up to a little over 100% because of the platforms’ spread, so we scale them back to a clean 100% split. Where the order book is tight we use the midpoint of the best bid and ask, and otherwise the last trade.
• Thin markets. Smaller races trade far less than chamber control. A market with under $10,000 of volume is labeled thinly traded, because a single bet can move it.
• An indicator, not a forecast. Market prices are context only. They are not used in any rating, seat count, scenario or the daily forecast, and they are not betting advice. Traders can be wrong, and prediction markets have moved with news and with the people trading them.

Daily update

• Each morning at 6am Pacific the site compares the current ratings and polls with a saved snapshot of the previous update.
• The headline for each chamber reports the change in Dem-leaning, Toss-up and Rep-leaning seats and in distance from majority, then the races whose rating moved and the polls and rater changes behind them.
• An early voting section opens with a plain words read of every connected state together. It says whether Democrats are ahead of, level with or behind 2024 and 2022 at the same point, how the Republican and other voter shares moved, how ballot volume compares, which states are moving most, and what changed since the last update. A second read covers the toss-ups in each chamber, with what it means and which toss-ups have no party data and why. The words are generated from the numbers by fixed rules, where a gap of one point or less counts as level, and they are context only and not a forecast.
• A betting markets section reports the chance Kalshi and Polymarket give each party in each chamber and across the toss-ups, with the change since the previous update.
• The 7-day trend only compares updates that used the same method version, so a change to the method doesn’t show up as a political shift.

Limits

• The ratings are other people’s judgments averaged together, so the forecast is only as current and independent as those raters, and several of them are themselves driven by polls and models.
• A race at 0.4 and one at 0.6 can land in different categories. The averages behind the ratings are shown on each race page.
• Polling averages here are simple, transparent weights, not a statistical model. They don’t estimate house effects or correlated error between races.
• The what-if scenarios assume 2026 error repeats 2024’s size and direction. Polling error has gone both ways in other years, and state results for 2024 come from few polls in some states.

Change log

Any change that moves the numbers is listed here with its method version.

• v2 update · 2026-10-20 Added betting-market odds from Kalshi and Polymarket for House, Senate and governor control and for each toss-up race, refreshed with the data and summarized in the daily update. This is context only and does not change any rating or seat count, so the method version is unchanged.
• v2 update · 2026-10-15 Added early and absentee voting by party for toss-up races, starting with Pennsylvania, North Carolina, Maine, Iowa, Florida and Nevada's Clark County, compared with 2022 and 2024 at the same point before Election Day with Florida as finished totals only, and a summary of it in the daily update. This is extra information and does not change any rating or seat count, so the method version is unchanged.
• v2 update · 2026-10-15 What-if scenarios only. Iowa's 2024 state polling error now leaves out the final Iowa poll from Selzer & Co., which had Harris +3 in a state Trump won by 13, so Iowa's error is 5.5 points, down from 8.5. The current forecast and daily update are unaffected, so the method version is unchanged.
• v2 · 2026-10-12 Tied averages now round to the less decisive rating on both sides, so a race averaging exactly +0.5 or -0.5 is a Toss-up. Poll averages ignore polls older than 14 days once a race has 5 or more polls in that window. Added the 2024 polling-error what-if scenarios.
• v1 · 2026-10-07 Launch with the consensus as the average of the raters, an independent-candidate rule, editorial overrides, and quality-weighted poll averages with a 14-day half-life.

Back to the forecast