Six ways of forecasting the 2026 midterms, collected every day and shown side by side. The display is the disagreement between methods, not another trend line.
Five families, each a group of forecasts that share a way of knowing. They are listed here in the order the site draws them, which runs from least modelled to most.
| Family | What its members have in common |
|---|---|
| Polling | Starts from what people tell pollsters, then averages, weights, or corrects for how early it is. |
| Markets | Prices from people betting money — an aggregate of everyone else's models plus whatever the traders think they know. |
| Fundamentals | Predicts from conditions rather than opinion: the economy, presidential approval, seats defended. Knows nothing about candidates or polls. |
| Professional | Published statistical forecasts combining most of the above plus candidate quality, fundraising and local knowledge. Shown as an average once enough of them are collected on terms permitting it; individual numbers appear only where a publisher allows. More. |
| Academic | Specifications published in the political-science forecasting literature, rebuilt here and run on our data. Published equations, run on our own inputs, which spread more widely than any other family. |
A forecast can belong to more than one family, and several do. The first four labels describe a METHOD — what the forecast looks at. “Academic” describes PROVENANCE — where it was published. Those are different questions, so a referendum model published in a journal is an academic forecast and a fundamentals one, and it counts in both averages.
Expert ratings are collected but kept out of every average: “Lean R” is not a number, and choosing which number to make it is a modelling decision. They have their own section and a graphic.
Everything in our two models that is a judgment rather than a measurement. The academic models carry their own, listed with each model in the academic section — Lockerbie's open-seat term is the clearest example, since its sign is set by hand rather than estimated.
| Choice | What we did | Why it is arguable |
|---|---|---|
| Generic-ballot shrinkage | λ decays from 1.0 at Election Day toward 0.5 far out; λ=0.91 today | A stylised reading of Bafumi–Erikson–Wlezien, not a re-estimation. |
| Race-level uncertainty | σ=5.4 points, from official Senate returns against prior-cycle lean | One cycle, n≈30. Cannot separate a national polling miss from state-level noise. |
| Seat baseline | 34 Democratic holdovers | Bookkeeping, not forecast — worth about twenty points of headline probability per seat. |
| Lean to margin | Expected margin = national tide + 0.92 × 2 × state lean | Linear, and it overshoots at the tails where there are not enough voters left to swing. The 0.92 is measured rather than assumed: a composite index says seats are about 8% further apart than they turn out to be. Three independent fits agree — 0.924 on Senate returns, 0.921 on 768 House district-cycles, and 0.902 implied by what two race-level forecasters separately think about the 2026 Senate map. Until 28 August 2026 the Senate model used 1.0 while the House model already used the fitted value, which is an inconsistency rather than a choice. Correcting it moves expected Senate seats by about half a seat and the probability of a Democratic majority by four to six points, because compression brings lean seats within reach of the national shock even though the competitive races barely move. The Senate fit rests on 30 state-years from one cycle; the House fit is what makes it credible. |
| Senate incumbency term | +5.70 margin points, from 28 August 2026 | Until 28 August 2026 the Senate model applied no incumbency term at all, because the archive held no roster of which party holds each of the 35 seats up or who is retiring. That was the largest known gap in the Senate numbers, and Maine showed its size: a Republican incumbent in a D+7.6 state that two race-level forecasters independently put about eleven points more Republican than lean alone implies. The roster now exists and the term is +5.70 points, fitted jointly with the lean slope and the spread, because pairing a no-incumbency slope with a with-incumbency spread would claim an accuracy the projection has not earned. |
| The Senate history spans two methods | Step on 28 August 2026 | Both corrections above — the fitted lean slope and the incumbency term — were applied going forward rather than backward through the archive, so the Senate lines carry a step on 28 August that no forecaster caused. It is visible and it is ours: the probability of a Democratic Senate majority on the polling line falls from 0.24 to 0.16 that day while every contributing tide is unchanged, because the same tide now maps to fewer seats. Every other family steps the same way on the same date for the same reason. The seat history is otherwise retrospective throughout — today’s model run on each past date’s evidence — so a reader is entitled to expect one method across it, and until the history is re-projected on the current calibration that expectation is not met. Dates before 28 August 2026 should be read as the older specification. |
| No bias correction | None applied | Generic-ballot bias is conditional on approval rather than constant. |
| Approval input | Adults-sample polls only | The two models that use approval were fitted on a Gallup series, and Gallup interviewed all adults. Likely-voter samples currently run about eight points better for this president than adult samples, so the population is the choice that matters. It puts our approval several points below every published tracker, which average all three populations together. |
| Two-party assumption | Every race modelled as D versus R | Nebraska has no Democrat and a serious independent; the model cannot represent that. |
| Which district map | The lines in force on the date being shown, not today’s | Ten states redrew mid-cycle. Where a map has no signature date — a court imposed it, or it took effect only when an injunction lifted — the date is a judgment call, and California’s is worth about three and a half seats. |
| Income in January | The last complete year-to-date figure is carried forward until the new year’s first month is published | Income enters both models as this year’s months averaged against last year’s, and on 1 January there are no months of the new year yet. With a two-month release lag the figure does not exist again until about 1 March. Carrying the last real value keeps the roster constant; the alternative was for two models to drop out and the family average to lurch every week. The cost is that these lines are flat by construction for the first two months of each year, and a reader cannot tell stability from an absence of data by looking. The provenance field on every affected row reads carried. |
Presidential approval, real income growth, and seats defended. Fit on all 20 midterms from 1946 to 2022 and scored by leave-one-election-out error.
adults only, what the models use aggregator average – – all polls, any population ◦ Gallup’s own readings, to Dec 2025
This model cannot. Its coefficients were fitted on a Gallup series running back to 1946, and Gallup interviewed all adults, so feeding it a mixed-population average would be a different model reported under the same name. We average the adults polls, and the gap you see against other trackers is that choice rather than a disagreement about the polls. Both numbers are in the table above.
Gallup stopped polling presidential approval in December 2025, but it published twelve readings first, and those twelve are the only real test of which construction stands in for it best. Scored on how closely each reproduces an actual Gallup number: the adults average misses by 2.1 points on average, a house-effect-adjusted whole-field average by 4.0, and the raw whole-field average by 5.2. All three are equally noisy; the difference is almost entirely a constant offset. That is the argument, and it is why the choice is not a matter of taste.
Gallup’s own remaining house effect, measured against the rest of the adults field, is about two points. We do not correct for it. The estimate drifts across those twelve months, an independent estimate from a published model puts it nearer three quarters of a point, and it can never be checked again now that Gallup has left the field. It stays in the error bar.
Its lineage, since people ask: the first two terms are the referendum specification that goes back to Tufte (1975) — a midterm as a verdict on the administration and the economy. The third is an exposure term, in the sense of Oppenheimer, Stimson and Waterman (1986): a party defending more seats has more to lose. It is not Abramowitz’s “Time for Change”, which is a presidential-year model with different inputs and a different dependent variable. The academic section runs the unconditioned referendum specification beside this one, and the gap between the two lines is the exposure term doing its work.
R two-party House vote = 24.49 + 0.134 × approval + 0.312 × real income growth + 0.069 × seats held going in
| Input | Value | Source |
|---|---|---|
| Presidential approval | 35% | 5 polls of all adults in the past fortnight, unadjusted |
| for comparison: the aggregators | 37.53% | Mean of 10 published approval trackers, spanning 35.0% to 40.0%. Not a model input. |
| for comparison: all polls, any population | 40.58% | The same polls with the population filter off. Not a model input. |
| Real income growth | -0.039% | FRED / BEA, year to date |
| Seats held going in | 220 | current chamber |
| Output | Today |
|---|---|
| National House margin | D+11.2 (80% 5.4 to 17.1) |
| Fit | LOO RMSE 2.29 pts · in-sample R² 0.66 |
| Approval | D margin |
|---|---|
| 31% | D+12.3 |
| 33% | D+11.8 |
| 35% | D+11.2 |
| 37% | D+10.7 |
| 39% | D+10.2 |
The seat counts come from pushing this margin through the same seat machinery the polling model uses, so any difference between the two models is a difference in the national tide and nothing else.
Three equations estimated together — presidential, on-term House and midterm House — on data back to 1916. Reads the economy directly from growth and inflation where ours reads it through approval, and carries an explicit balancing term, so a party that has done well is predicted to lose ground at the following midterm. Forecast page · paper
We take the number Fair publishes rather than re-estimating his equation. The equation is in the paper and is reproducible; its inputs are not — the inflation and growth terms for a cycle still running come from his own quarterly macro forecasts, which are a separate model we do not have. Recomputing it with our guesses at those would produce a number that is not his forecast while carrying his name. His seat projection here is ours: his national margin pushed through the same partisan lean as every other model on this page, so the two are comparable.
FRED supplies real disposable personal income per capita, pulled daily. MIT Election Lab supplies certified official returns under a public-domain dedication — the basis for our state partisan lean, and the reason the Senate table can be published at all.
The two chambers use different baselines, and that is deliberate. House districts are leaned on Dave’s Redistricting App, whose composite we re-centre on the mean of the 435 current districts so it means the same thing Cook’s index does. States are leaned on our own reconstruction from the MIT returns above. The reason is availability rather than preference: DRA computes a partisan composite per district and has no consistent statewide product, so a Senate baseline built from it would have to be reassembled from districts with vote weights we do not hold for every state, and the six single-district states would still have to be filled from the MIT composite regardless. Because the slope, the incumbency coefficient and the three sigmas are all properties of the index they were fitted against, each baseline carries its own calibration and the two are never mixed. Every projection records which baseline produced it.
Specifications published in the political-science forecasting literature, rebuilt here and run on our data.
These are our implementations, not the authors’ own forecasts, and the distinction is not a formality. Where an author has published a 2026 number we say so and print theirs beside ours; where we have changed the specification to fit the data we hold, we say that too. A model here wearing somebody’s name means we followed their equation, not that they endorsed the result.
Why this family exists at all: an academic model publishes an equation, where a commercial forecaster publishes a number. Rebuilding an equation and running it on inputs we captured ourselves is ordinary scholarship, so nothing here waits on a licence.
Bafumi, Erikson & Wlezien, 'Balancing, Generic Polls and Midterm Congressional Elections', Journal of Politics 72(3), 2010
Our implementation of BEW's published midterm equations, run on our generic-ballot average. Uses the authors’ published coefficients, unchanged.
The paper estimates a seats equation separately from its vote equation, and the two do not agree once applied to 2026: theirs returns 244 Democratic seats directly, while their vote number pushed through our district baselines returns rather fewer. The seats equation was fitted on 1946–2002 and carries the mid-century House’s seat bonus, which current districts do not give. We use the vote equation, so this model is comparable with every other line on the site, and leave the gap on show.
tide is BEW's vote equation; their separate seats equation is reported as a diagnostic and is not used downstream
Lockerbie, 'The Challenge of Forecasting the 2024 Presidential and House Elections: Economic Pessimism and Election Outcomes', PS: Political Science & Politics, 2024
Our implementation of Lockerbie's published House equation. The open-seat term's SIGN is our assumption, not his — see the sensitivity table. Uses the authors’ published coefficients, unchanged.
Coefficients: constant 5.360; open_seats_signed 0.410; out_of_sample_mae_seats 16.800; pct_expect_worse -0.780; published_2024_forecast -12.000; r2 0.420; r2_adj 0.380;
The month is part of the specification. Lockerbie reads June of the election year, and June 2026 came in below the months around it. We feed it June anyway: swapping in a smoothed average would be estimating a different model and reporting it under his name. Here is what the choice is worth.
| Reading | % expecting worse | President’s party seat change |
|---|---|---|
| Mar 2026 | 39% | -50 |
| Apr 2026 | 41% | -51 |
| May 2026 | 45% | -54 |
| Jun 2026 · the specification | 37% | -48 |
What the open-seat term is worth. This model multiplies open seats by the direction of the year. We first read that as a judgment we would have to make ourselves — but the paper settles it: “Midterms are, by definition, a bad year for the incumbent party.” The −1 is the author’s rule, fixed for every midterm, not our call about 2026. The other rows are counterfactuals, shown because this one term carries about half the forecast:
| If 2026 is… | President’s party seat change | Implied D seats |
|---|---|---|
| a bad year for the president's party — the midterm rule | -48 | 263 |
| a neutral year | -24 | 238 |
| a good year for the president's party | +1 | 214 |
Any model of this kind carries judgment calls — a shrinkage curve, a winsorised year, the width of a window. This model puts one in the open as a variable with a sign, which is why it is worth pausing on even though the midterm case turns out to be specified rather than chosen. In an on-year election the author calls it “a little more complicated”, and there the judgment really would be a judgment. The highlighted row is the one we publish.
Read the point estimate with its error bar and not without it. The author publishes an out-of-sample mean absolute error of 16.8 seats across 35 elections, and individual years have missed by more. The interval is his own, and it is wide by design. The interval on the figure above is built from that published error, which is why it is so wide.
forecasts SEAT CHANGE directly — the margin shown is our own seat curve run backwards from that seat count, and is not a number Lockerbie publishes
that seat count falls outside the range of tides the other models produced today, so the inversion is extrapolated rather than interpolated
OUTSIDE THE FITTED RANGE on pct_expect_worse = 37, above the sample maximum of 32 and open_seats = 60, above the sample maximum of 56. The model was fit on 1952-2022 and this is a linear extrapolation beyond anything in that sample, which is most of why it is the most Democratic forecast on this site.
Lockerbie: 'Midterms are, by definition, a bad year for the incumbent party.' The -1 is his rule, not our assumption.
published out-of-sample mean absolute error is 16.8 seats — the interval is his, not ours, and it is very wide
Lewis-Beck & Quinlan, 'A Political History Forecast of the 2024 US Congressional Elections', PS: Political Science & Politics 58(2), 2025, Table 1
Our implementation of their published Table 1 equations. Forecasts both chambers directly. Uses the authors’ published coefficients, unchanged.
The authors’ own published figure for 2024: Democrats lose Senate control with a net loss of three seats; House a knife-edge race at 215 D seats (Lewis-Beck & Quinlan). their 2024 forecast, quoted so our implementation can be sanity-checked against the published one
forecasts SEAT COUNTS directly from institutional facts — no polls, no economy, no popularity. The margin shown is our seat curve run backwards from its House number, not theirs
its Senate figure of 49 D seats is computed from a SEPARATE published equation, not from our partisan-lean machinery — the only Senate number on this site that is independent of it
Tufte, 'Determinants of the Outcomes of Midterm Congressional Elections', APSR 69(3), 1975; the modern version is Lewis-Beck & Tien
Our fit of the referendum specification. NOT Lewis-Beck & Tien's published forecast. Refit on our twenty-midterm history rather than using the authors’ published coefficients.
Coefficients: approval 0.026; income_growth 0.683; intercept 44.938; R² 0.27, leave-one-out RMSE 3.18 points of vote share.
The authors’ own published figure for 2026: Republicans lose about 28 House seats and the House (Charles Tien and Michael S. Lewis-Beck, LSE USAPP blog, 2025-10-13). their DV is seat change, ours is vote share — the comparison is directional, not a reproduction
no exposure term — this is the unconditioned referendum specification, and the gap against our fundamentals model IS the finding
Models we would like in this family and have not implemented. Listed because a methods page that shows only what we managed to build reads as a complete account of the literature.
State approval + state economy (Enns et al.)
Enns, Lagodny, Colner & Kumar, 'Understanding Biden's Exit and the 2024 Election: The State Presidential Approval-State Economy Model', PS: Political Science & Politics, October 2024
The generic ballot as a national tide, shrunk toward even because early leads overstate the eventual result, then carried to each state by its partisan lean.
| Step | Value |
|---|---|
| Generic ballot average | D+6.74 |
| Shrinkage λ | 0.906 |
| Nowcast tide | D+6.74 |
| …shrunk toward November (diagnostic, not used) | D+6.11 |
| Race-level σ | 5.37 pts |
| D Senate seats | 49 of 100 (80% 47–51) |
| P(D reach 50+) | 42% |
| P(D majority, 51+) | 20% |
| Holdover D seats | P(D reach 50+) |
|---|---|
| 33 | 20% |
| 34 (used) | 42% |
| 35 | 74% |
The model does not run its own polling average; it takes one, which makes the choice of whose a modelling decision.
| Source | What it is |
|---|---|
| Silver Bulletin | Generic-ballot average with house-effect adjustments and recency weighting. The series this model reads. |
| RealClearPolling | Simpler unweighted average of recent polls. Not collected directly: their site is not open to automated collection. Their average reaches us only where a third party republishes it with attribution. |
| Race to the WH | Per-race poll list, collected for the Senate. |
| Wikipedia's aggregator table | A CC BY-SA table listing several aggregators' current generic-ballot averages side by side — Decision Desk HQ, RealClearPolling, Silver Bulletin, Race to the WH and others. Each row is attributed to the aggregator that produced it and inherits that aggregator's own licence, so nothing is republished that its owner has not permitted. It is a substantial part of what the polling average is built from. |
Silver Bulletin and Race to the WH permit collection but not per-poll republication during the cycle. Note what the polling category average actually mixes: most aggregators publish TODAY's generic ballot, while our own model contributes its ELECTION-DAY tide, which is that number shrunk toward the eventual result. They are both polling-based estimates of the national House margin, but they are not estimates of the same instant.
Published models combining polling, fundamentals, candidate quality and money. The category we are least free to reproduce.
So the professional line is absent, not late. Collection continues daily and every row is kept; the full series will be in the post-election release, when the licence questions that govern the cycle no longer apply.
The price is that a seat count shown against a forecaster’s name is our model reading their generic ballot, not their own seat forecast, and the two can disagree. Race to the WH is the clearest case: it publishes thirty-five Senate race margins of its own, and none of them enter the Senate number shown here. Those race-level figures are collected and do feed the race-level averages; it is only the chamber totals that are reconstructed. Where a forecaster publishes its own seat count, read theirs rather than ours.
One visible consequence: three aggregators have published the same rounded generic ballot for several days running, so their seat counts and chamber probabilities are identical to the last decimal. That is the method working rather than a duplicate row. It also means the number beside a family is not as many independent readings as it appears, which is what the contributor count beside it is for.
| Forecaster | What it is |
|---|---|
| Grant Williams | An open-source district- and state-level model built on partisan lean, polling, incumbency and candidate quality. MIT-licensed. Collected daily. |
| Decision Desk HQ | Polls-plus-fundamentals model paired with a live results operation. Their forecast page is not collected — not at their request, which we wrongly said here until 29 August 2026, but because the licence question is open and we have not yet asked them. Their robots.txt permits the path. Their race ratings reach us, attributed, where a third party republishes them. Worth knowing if it is ever enabled: their model ingests Polymarket and Kalshi by volume weight, so it is downstream of the markets already collected here and could not be averaged beside them without counting those twice. |
| The Economist | Bayesian model pooling polls across states with a fundamentals prior. Subscriber-only and not collected. Their race ratings reach us, attributed, where a third party republishes them. |
| Race to the WH | Race-by-race probabilities with a long published history of its own trend lines. |
| Split Ticket | Candidate-quality metrics — how far a nominee runs ahead of or behind the seat's baseline. Now published at The Argument. Registered but not collected. |
| FiftyPlusOne | G. Elliott Morris, successor to earlier work at The Economist and FiveThirtyEight. |
| VoteHub | Polling aggregation and race ratings. |
What we publish for this category is the average across everyone in it — wisdom of the crowd rather than a pick — and it is on the comparisons page. Most of these forecasters permit collection but not per-forecaster republication during the cycle anyway; see what is not published.
Not models. A price is what someone will pay for a contract that pays out if an event happens; reading it as a probability is a good approximation rather than an identity — fees, the cost of tying up capital until November, and a documented tendency to overprice longshots all push it around at the extremes.
| Market | What it is | Today |
|---|---|---|
| Polymarket | Real-money markets on chamber control and individual races, plus a national popular-vote margin ladder that resolves both tails. | 52% D Senate 88% D House D+7.7 margin |
| Kalshi | CFTC-regulated US event exchange running the same contracts, and the only one quoting a ladder over seat counts — which is where the market seat projection comes from. | 50% D Senate 231 D House seats 50 D Senate seats D+7.7 margin |
| PredictIt | Real-money exchange with an $850 position cap and the longest academic record of the three. Alone among them it lists individual Senate races, which is where the market column on the comparisons page comes from. | 48% D Senate 84% D House |
A price here is the LAST TRADE, which on a thin market can be hours old and on either side of a wide spread. Nothing smooths it, and no market on this page is adjusted for longshot bias.
Some of these contracts are not a price but a ladder: a row of markets, each paying out if the final number lands in one bucket. Kalshi runs one over Democratic seat counts (“218–221”, “Above 249”), and Kalshi and Polymarket both run one over the national House popular-vote margin (“Democrats, 8 to 10%”). A ladder is a probability distribution rather than a single number, which is more than any other source on this site provides, and it is where the market seat count and the market vote margin come from. Without it the markets column would hold chamber-control probabilities and nothing else — a control price does not imply a seat count, and it certainly does not imply a vote share.
Two conventions turn a ladder into a number, and both are choices rather than arithmetic. Prices are bid/ask midpoints across a dozen contracts, each carrying its own spread, so they do not sum to one; everything is normalised by their sum rather than trusted as a distribution. And an open-ended bucket has no honest representative value — “Above 249” is a claim about everything up to 435 — so it is read as one more bucket of its neighbours’ width. That understates the tails on purpose. The alternative, using the chamber maximum, would have “Above 249” imply an average of 342 seats.
PredictIt’s seat-count market uses wide buckets (“192 or fewer”), too wide to read as a distribution, so it is not parsed as one. Its contribution here is chamber and per-race probabilities.
Ordered categories — Safe, Likely, Lean, Tilt, Toss-up — produced by people who talk to campaigns.
11 raters, transcribed from Wikipedia under CC BY-SA 4.0 with attribution: Cook Political Report, Decision Desk HQ, FiftyPlusOne, Fox News Power Rankings, Inside Elections, Race to the WH, RealClearPolitics, Sabato's Crystal Ball, Silver Bulletin, Split Ticket, The Economist.
They are kept out of every average. Turning “Lean R” into a margin requires picking a number, and whichever number you pick is a modelling assumption doing real work while looking like data entry. What ratings are good for is disagreement, which is drawn as a spread on the comparisons page.
The rules were fixed on 22 August 2026, seventy-three days before the election and before any 2026 result existed. They are in the repository as forecast/score/RULES.md, where git records when they appeared and every change since.
Writing them afterwards would have been worth much less. Which horizon counts, what “control” means, whether unopposed races belong in the denominator — each of those changes who comes out ahead, and a rule chosen after the answer is known cannot be told apart from a rule chosen because the answer is known.
| Question | Answer fixed in advance |
|---|---|
| Who is scored | Every source and every category average, by the same code. Our own class models get no exemption. |
| When | Nine fixed horizons — 180, 120, 90, 60, 30, 14, 7 and 1 day out, plus the last value before polls open. A forecast staler than 21 days at a horizon is recorded absent rather than carried forward or penalised. |
| On what | Brier score on chamber control, absolute and signed error on the national margin and on seats, race-level Brier with a calibration table, and a log score on the seat ladders where an exchange quotes a distribution. |
| Senate control | 51 seats for the Democrats, not 50. The Vice President breaks ties through this Congress and is a Republican, so 50–50 is Republican control. |
| The national margin | Two-party, over all 435 districts, votes as cast, unopposed races included. The version excluding unopposed races is published beside it as a robustness check, never substituted for it. |
| Truth | Certified returns as they stand on 6 January 2027, after any runoff state law requires. Stored in the archive with its hash, so the numbers that resolved the scores can be checked rather than taken on trust. |
| Against what | Four naive baselines, scored by the same code: a coin flip, the postwar base rate, the 2024 result carried forward, and the final polling average. The no-change baseline is what every other entry is read against. |
Ten states redrew their congressional maps during this cycle and 123 of the 435 districts moved, the last of them on 2 June 2026. That breaks something that looks like bookkeeping and is not. A national margin does not care where the lines are, but a seat count does, and every seat count on this site before today was originally produced by pushing a past national tide through today’s districts — districts that, for most of 2025, did not exist.
They are now computed on the map that was in force on their own date. The baseline is assembled per state rather than chosen wholesale: on 1 December 2025 that means Texas, Missouri, North Carolina and Ohio on new lines with California, Florida, Tennessee, Louisiana, Alabama and Utah still on old ones. The correction is worth about six Democratic seats at the start of the archive and falls to nothing by June 2026, so the current numbers are unchanged and the history is what moved. Each projection records which baseline produced it.
The effective dates are published with the basis for each, because five of the ten cannot use the obvious rule. North Carolina’s map became law without a signature, Utah’s was imposed by a court, Alabama’s was passed conditionally and took force only when the Supreme Court lifted an injunction, and Ohio’s date comes from our own archive of the Wikipedia status table. California signed on 21 August 2025 but the map could not govern a ballot until voters ratified it on 4 November, and the ratification date is used. That single choice is worth 3.6 expected seats across the intervening seventy-five days, which is why it is written down rather than left in a comment.
What this looks like on the chart is worth saying plainly, because it is easy to misread. Every line steps on each of the eleven dates a state’s new map took effect, and they step together, because the ground moved rather than anyone’s forecast. Florida on 4 May 2026 is the largest: about four and a half seats off every model on the same day, including Ray Fair’s, whose published vote share did not move at all that week. A step that appears in one line is news. A step that appears in all of them at once is the map.
Two consequences for scoring, both fixed before any result exists. A race-level forecast is scored only if it was made under the map that decided the race, since a forecast of TX-09 in March 2025 is about different ground from the TX-09 on the November ballot. And a backfilled seat count computed on the right map is still a number we produced later, from a specification chosen later, so it stays out of the real-time table. The fix makes those rows accurate. It does not make them evidence about what was knowable at the time.
Nothing is scored yet, because nothing has happened yet. What runs today is the coverage report: which forecasters will have a value at each horizon, so a gap can be closed while there is still time. It has already found one — most named forecasters were not being captured before 19 August 2026, so the three earliest horizons carry the category averages and the sources whose history could be reconstructed, which now includes both of our own models.
This archive is maintained by the same person who wrote two of the models being scored. The protections are that the rules were fixed in advance, that the code path is identical for every source, and that the raw archive is released afterwards so anyone can recompute every number from the stored bytes. That is not independence and is not offered as such.
Those appear only inside an average with at least three gated contributors, so no individual forecast can be backed out by subtraction. The floor counts gated sources rather than all sources: a source we publish by name adds nothing to anyone's protection, since a reader can subtract it straight back out. Enforced in code, and the aggregator re-derives it from its own output and refuses to write if it is violated. Today it withholds 51302 cells.
District forecasts are built from a licensed partisan index, and because we publish the national tide, a district margin and that index are recoverable from one another. So no district margins are published. Chamber totals are: a seat count is a sum over 435 districts and cannot be inverted back to any of them. The Senate has no such problem — its state lean is our own reconstruction from public-domain returns.
Category averages, our own models in full, the individual forecasts whose licences permit it, and a manifest of SHA-256 hashes for every capture, committed on the day it was taken. When the full archive is released after the election, anyone can verify the released bytes are bit-identical to what was collected on the dates claimed.