The Margin · Methods
Methodology
How The Margin turns the public record into projections, in plain English. Everything here is a nowcast, what current polling implies if an election were held now, not a prediction of a future result. The house philosophy is honesty over coverage. We publish what the data can support, disclose every simplification, and leave a number out rather than invent one.
01 · Every political page
Poll averages
Every political page starts from the same average. Recent published national polls are blended so newer fieldwork counts for more, and the headline figure is always the blend, never a single poll.
Inputs
Published national polls from named polling houses, compiled from the public polling record and dated by fieldwork. The small print on each page lists exactly which pollsters are in the blend, and the raw polls behind every average are downloadable as CSV from the data page.
Engine
A recency-weighted mean with an exponential 10 day half life, so a poll two weeks old carries well under half the weight of one from today, with a light pull toward larger samples.
Uncertainty
The ± figure is a 95 percent between-poll band. It measures how much the recent polls disagree with one another, scaled by the effective number of polls, with a floor of ±1.1 points. It does not capture every source of survey error.
Limits
Sparse feeds are shown, not hidden. Where a country has only one or two recent polls the average is still published, with its staleness flagged, rather than dressed up as more than it is.
02 · One engine per system
Seat models, system by system
The poll average is pushed through each country's real electoral rules, one engine per system. Where a system cannot honestly be reduced to a seat count, we decline to publish one.
First past the post
United Kingdom and Canada use uniform national swing. Each constituency's result at the last election, 2024 across all 650 UK constituencies and the official 2025 result across 343 Canadian ridings, is shifted by the change in each party's national share, and the winner of every seat is called. Transparent, fast and the standard first-order model, though it does not capture local candidates or region-specific trends.
Proportional
Under the 2025 reform the Bundestag is allocated strictly in proportion to party second votes. We apply the 5 percent threshold, then distribute all 630 seats by the Sainte-Laguë method. The constituency map swings each Wahlkreis's 2025 second vote by the national movement to show the leading party locally.
List PR
For the other European chambers we apply each country's legal threshold to the national poll shares and allocate seats by the divisor rule the country actually uses, D'Hondt or Sainte-Laguë, as a single national pool. District structure, levelling seats, STV transfers and majority bonuses are simplified away, and each country page states exactly what its approximation omits.
Mixed
Italy's Rosatellum. Roughly a third of seats are single-member districts awarded to the leading coalition, and the rest are proportional to party lists above 3 percent, allocated by largest remainder. The district map swings each 2022 collegio by the national coalition movement.
Two rounds
France elects deputies over two rounds with tactical withdrawals, so a national poll cannot be turned into a reliable seat count. We show the projected first-round leader in every circonscription, its 2024 first round swung by the national movement, and do not put a single seat number on it.
Preferential
Australia's House tracks the two-party-preferred vote. Every classic seat is swung by the national 2PP movement off the 2025 pendulum, and the crossbench, the teal independents and the Greens, is held constant because a 2PP poll does not measure it.
Europarl
The European Parliament projection maps national poll averages to European political groups and allocates each member state's MEP quota by largest remainder, summing to exactly 720 seats. It is a deliberate approximation of the group balance, not a forecast of any single seat.
03 · The deep model
The US midterm model
The 2026 House and Senate forecast is the site's deepest model and the only engine that goes beyond a pure polling nowcast. Like every figure here, it describes November as it looks today, not as it will be.
Inputs
The national generic congressional ballot average, the share saying they would vote Democratic or Republican, plus the full 2024 results baseline, real special election results and a deliberately small dose of individual race polls.
Environment
The national environment is estimated as a swing against the 2024 result, never from raw poll levels. It blends the change in the generic ballot since 2024, a special election overperformance index built from actual votes, and a historical midterm penalty against the president's party.
Seats
Each race's forecast margin starts from a blended 2024 baseline, then adds the seat's elasticity times the national swing, incumbency, net candidate quality and a low-weight race-poll adjustment.
Simulation
25,000 Monte Carlo runs with correlated error at the national, regional, demographic and individual seat level, so the chamber probabilities respect the ways polls have actually missed before.
Ratings
Alongside the model, the states page keeps a consensus-style ratings board for every 2026 Senate and Governor race, from Safe to Toss-up. It is a curated snapshot of the public ratings consensus, not a model output, and is labelled as such everywhere.
Limits
Governorships carry ratings only. Where the consensus on a race is unsettled it is marked uncertain rather than forced onto the ladder.
04 · Below the national level
Subnational and devolved
Below the national level the site prefers verified baselines to thin modelling.
Coverage
Every Canadian provincial legislature, every Australian state parliament, and Scotland, Wales, Northern Ireland and London on their own devolved pages.
Engine
Standings are shown as last elected, a clearly labelled baseline built from official results. Where credible provincial or state polling exists, for example Ontario, British Columbia, Alberta and Queensland, a uniform-swing estimate off the last result is layered on top and the pollsters are named.
Uncertainty
Baseline pages carry no uncertainty claim because they are results, not estimates. Polled pages inherit the caveats of small-sample regional polling.
Limits
Live polling is added as it is published. Until then a chamber shows its composition, not a projection, and says so on the page.
05 · The same treatment
Sport
The same numbers-first treatment, with engines matched to what the data can support.
Elo
MLB and NFL power ratings are Elo, built from scratch off game results on the logistic 400-point curve with home advantage and a margin-of-victory multiplier. MLB is rated live from this season's games. NFL carries ratings from the last completed season, mean-reverted by one third.
Playoff odds
A Monte Carlo simulator plays out the remaining schedule thousands of times from each matchup's Elo win probability and counts playoff, division, pennant and title outcomes.
World Cup
The 2026 tournament is simulated from each surviving team's real bracket position, 20,000 continuations on the same Elo curve, seeded from the public World Football Elo record. Matches already played are never re-simulated.
Brackets
The Formula 1, NBA and NHL pages are season records, standings, playoff brackets and calendars built from verified results and cross-checked before publication. No odds are attached to them.
Limits
Elo sees results, not rosters. Injuries, trades and lineup news are invisible to the ratings until they show up in the games.
06 · How it runs
Data pipeline
Every number on the site is produced by an automated, inspectable pipeline and published back out as open data.
Refresh
Scheduled jobs refresh the poll feeds and the US midterm model daily, and the sport models nightly in season. Each page shows when its data was last updated.
Geometry
Constituency maps are drawn from vendored official boundary files, among them Elections Canada's federal districts, the AEC national set, the Bundeswahlleiterin Wahlkreis file, the Ministero dell'Interno collegi and the data.gouv.fr circonscriptions. A map only exists where an official source does.
Sources
Seat baselines come from national electoral authorities, and figures are checked against the original release. The full list is on the sources page.
Output
Everything shown is available as JSON and CSV from the data page, free to reuse with attribution.
07 · Disclosed limitations
What we do not model
Projections assume an election held today under current polling. No campaign, no events, no turnout model. They move as the polls move, and the numbers are estimates with real uncertainty, not destiny. In particular, none of the following is modelled.
- Turnout. There is no turnout model. Projections assume current polling translates into votes.
- Campaigns and events. A nowcast carries no view of what happens between now and polling day.
- Local candidates. Uniform swing treats every seat alike. Personal votes and local campaigns are not captured.
- Vote transfers. STV preferences in Ireland and Malta are not simulated. Those allocations are proportional approximations and are flagged as such.
- Run-offs. France's second round and its tactical withdrawals, which is why no French seat count is published.
- Majority bonuses. Greece's sliding winner's bonus is not modelled, so the leading party there is understated and the page says so.
- Coalition formation. Seats are projected. Governments are not.
- Sport news. Injuries, trades and suspensions reach the ratings only through results.
Every approximation above is also disclosed in place, on the page where it applies. Material changes to methods and figures are logged in the open on the corrections page. Sources are cited on each country page, and the model and data are open, so if a number looks wrong you can check it. Spotted an error? It matters to us. Write to balian.zane@gmail.com.