Call the Map is an independent, model-based forecast of the 2026 U.S. midterms: every House seat, every Senate race and every governorship, from one model. Free to read, no advertising, no account.
A forecast that shows its work. Every probability comes out of a model rather than a panel; every input that reached it is listed on the race page; and when a number moves, the site says whether the race moved or the model did.
Call the Map is not affiliated with, authorized by or paid for by any news organization, candidate, campaign, party or committee.
Every page, with no reader account, no advertising, no paywall and no data sold. Support is what keeps that possible, not what unlocks it. Running costs are ordinary and constant: hosting, map tiles, polling and finance data, the domain, and the machine time to fit, validate and reseal the forecast every day.
Those costs are the honest reason this section exists. This has been carried on one person’s time and one person’s money, and it is now at the point where it cannot improve much further without help. Anyone who wants to keep it going and make it better is genuinely appreciated.
If you contribute, please say in the message whether you would like to be listed — and if so, exactly what name to use, and whether you are happy for the amount to be shown. Say nothing and you stay anonymous, which is the default.
The received total is owner-reconciled from completed Ko-fi and Buy Me a Coffee provider records. The compute figure is a measurement at public list rates, not an invoice and not an out-of-pocket claim. It excludes hosting, data and subscriptions, the domain and my own time. The donation total is public because asking for support without showing what arrived felt a bit one-sided.
Ko-fi takes the smallest cut. You leave Call the Map to pay; the provider shows you its fees and privacy terms first.
It keeps the site online, and it lets me keep working on it. That is the whole of it. Hosting, map tiles, data, the domain and the machine time are ordinary running costs, and they arrive every month whether or not anyone contributes.
Beyond that it buys time — the thing this needs most. Every hour that does not have to be earned elsewhere is an hour that goes into more races, better tools, and eventually something more ambitious than one midterm. I would like this to still be here in 2028, and larger than it is now.
And you have my genuine thanks, which is not much of a return but is offered honestly.
No supporter tier of the forecast, no early access, no private numbers. A contribution never changes a forecast, a rating, a correction decision or editorial coverage.
Settled: for life. If a paid tier ever launches, anyone who supported Call the Map before it existed is never charged for it — not for the first year, but for as long as the site runs.
It applies to people who support the project before any paid tier is announced, and this published promise remains the record if the product changes later. Support is voluntary and does not purchase a subscription or guarantee a feature; do not assume it is tax-deductible. No card or bank details are collected on this site. For a payment question, write to hello@callthemap.com with the provider and transaction reference.
The roll call is taking its time.
No public names yet. The first supporters will enjoy excellent seating and very little competition for alphabetical order.
Early supporters may opt in to have their name listed here. Nobody is published by default, donation amounts stay private, and a contribution never changes the model or the coverage.
The 435 U.S. House districts, the 35 Senate elections and the 36 governorships on the ballot on November 3, 2026. For each race: a projected margin, a chance of winning and the middle 80% of outcomes. For each chamber: the chance of control, the expected seat count and the paths to a majority.
It does not forecast presidential politics, primaries or referendums. State-legislative and statewide down-ballot forecasts were withdrawn in September 2026 because too many of those races lacked usable prior history, confirmed candidate fields and district-level polling; they are outside the current public forecast, and bringing them back on evidence that can carry them is the first item on the list.
CTMM 2.2is one staged model: fundamentals fitted per office, then one national environment, then what the race itself knows — polls corrected for each pollster’s record, and the candidate field. Expert ratings are not model inputs. Finally all 506 races are simulated together, 100,000 times, with shared national, regional and state shocks, so chamber odds and race odds are one object rather than two calculations.
Missing evidence stays missing: a race with no polls is not given an imaginary tied poll. Campaign finance and prediction-market prices are collected, shown, and deliberately not used — the backtest switched finance off in every office, and market prices carry a weight of exactly zero.
One number, shared by every race, estimated from generic-ballot polling. Presidential approval and special-election results are not inputs: approval added nothing to a six-cycle regression, and special elections have no origin-dated history to fit against. Both are still collected and shown. Inspect the current environment →
Races are simulated jointly, so a national swing moves them together. The Scenario Machine reads pre-computed hypothetical boards; it never reruns the model or overwrites the published forecast.
Two documents describe this edition: a plain-language guide (PDF) and the technical methodology (PDF).
A chance of winning is a statement about a distribution, not a prediction. A candidate at 75 in 100 should lose about one time in four; when that happens the forecast is working. What you can hold it to is calibration: across many races, the 75s should win about three-quarters of the time.
Map fills show each race’s rating. Diagonal stripes mark a seat projected to change party; gray is no 2026 election, green an independent. On the closest-races lists the dot is the forecast margin and the band is the middle 80%.
CTMM 2.2was tested out of sample on the 2018–2024 record before it published a single 2026 number; the measurements are in the methodology's validation section. Four things to hold against it:
A public scorecard for 2026 — the exact forecasts issued here against certified results — is published after November 3.
Every forecast is published as an edition: a sealed bundle of inputs, rules and outputs, named by the hash of its contents. A source refresh never rewrites an older forecast; a change to a published number requires a new edition. This page reads the published edition ctm-2.2-20260913-ffd390c84f25, activated 2026-09-13T22:47:41+00:00.
Where a published number turns out to be wrong, the correction is written up on Substack rather than kept in a list here. The changelog remains the dated record of every material change, and each edition above is named by the hash of its own contents, so nothing published can be edited after the fact.
Call the Map is an independent project by one person: a political science student who previously worked in tech and has a background in web design — not a forecaster, a statistician or a journalist by profession. The method is published in full so it can stand on its own record rather than on anyone’s authority.
No party, campaign, donor, advertiser or prediction market can buy its way into the model. 2026 is the first cycle; the plan runs through 2028.
It started as a hobby project and is still largely one. It has taken far longer and cost far more than expected, and it was built by one person trying very hard to get it right rather than to get it out.
It does not claim to be better than anyone else’s forecast, or close to it. The people who have been doing this for years have earned their reputations, and no claim here is a comparison to them. What this site claims is narrower and checkable: that its method is published in full, that its numbers come from that method, and that where it is weak it says so on the page. Suggestions and corrections are always welcome — being corrected is better than being wrong in public.
The judgments are the operator’s: what counts as evidence, which trials ran, which results were accepted or refused, what ships. AI assistants did the parts around that — interface work, implementing and validating a specified method, acquiring and cleaning data.
What they never do is decide a forecast. Specifications are written before they are fitted, gates are fixed before results are seen, and a specification that fails its gates is refused rather than adjusted.
Work intended, in rough order — not a schedule, and not a list of things anyone is owed. Whether any of it happens comes down to how much time there is.
The forecast already refreshes automatically four times a day, each refresh sealing its own edition. The work now is speed and reach: new polls collected sooner, more picked up in a single pass, less time between a filing appearing and the board moving.
The whole site should work on an old phone on a bad connection, and today some of it only nearly does — the map and the largest race pages first.
The method is published in full, equations included, but a technical document is not an explanation. I want a clear trail from any number back to the evidence that produced it, in language that assumes no statistics background.
Readers will find things I missed. A reported error is reproduced against the edition it appeared in, fixed in a new sealed edition, and listed on the changelog.
The original plan was every race on the ballot, including 5,743 state-legislative seats. Too many of them have no usable prior history, no confirmed candidate field and no district-level polling. Bringing them back on evidence that can carry them is the thing I most want to do.
A results service of my own for the night itself: attributed returns, timestamps and reporting completeness beside the forecast they are testing.
Sourced analysis of individual races and reporting on the contests nobody covers. Writing stays separate from the numbers: nothing I write becomes a model input.
After November 3, the exact forecasts this site issued are scored against certified results and published, good or bad. Before then, the frozen end-to-end replay that would let this site quote an accuracy figure at all.
Longer pieces — what moved, why, and what the model makes of it — go out on Substack. The forecast itself lives here and updates on its own schedule.
Corrections, reproducibility problems and methodology challenges reach a person: corrections@callthemap.com, with the race and the edition concerned. Bug reports, interface complaints and methodological objections are all welcome.