Forecasting the House

The test

Four sources are combined. The polls come from six private outfits that each average the generic ballot, a survey question asking which party’s candidate you would support in your district, without naming anybody. Political scientists have shown that this one question, asked months out, predicts the national House vote well enough to forecast seats from.¹ No agency publishes such an average and none of the six publishes a readable file. They are read instead from the table that puts them side by side on the Wikipedia article “2026 United States House of Representatives elections” (https://en.wikipedia.org/wiki/2026_United_States_House_of_Representatives_elections).

The returns — the vote counts in every district — come from the Clerk of the House’s Statistics of the Congressional Election for 2004–2024, with the presidential vote inside each district beside them. The base rate, how the president’s party has fared at past midterms with nothing else known, comes from the House Historian’s record of party divisions back to 1858. The Senate roster, which senator holds which seat and when it is next up, comes from the congress-legislators file kept by volunteers on GitHub, and a published professional forecast is quoted for comparison.

Those sources answer the question because a forecast is a join between a national number and a local one, and both halves exist here in files with signatures on them. The key is the district, a state plus a district number, carried by the returns and the presidential figures alike. Every join takes on the assumptions of both files plus one more about how they line up, which is the subject of Join Keys and Crosswalks. That chapter also covers the crosswalk, a table that translates one set of areas into another.

The data

Twenty tables come out of this chapter’s build: the House forecast, the Senate forecast built the same way, and the joint-probability tables that put the two chambers together. Sizes are rows by columns, read off the files.

One more file is not data about the forecast – it is how this chapter checked its own work: checks.csv, the arithmetic verified before anything was drawn.

generic_ballot.csv holds one row per aggregator, with dem_pct, rep_pct, other_pct, margin and two_party_dem.

The six agree to within 0.5 points, and that is the last time in this brief that anything is this well determined. The last column is what the model uses. Between a tenth and an eighth of respondents pick neither party, and that share differs across the six by 4.9 points. Dropping them and recomputing gives the two-party share, averaging 53.73% Democratic.

baseline.csv holds one row per district. It carries district_id, the 2024 Democratic share of the two-party House vote in dv, the presidential share in dpres, the starting share the model uses in base, and the flags imputed, no_totals and uncontested, which mark a starting share that was estimated rather than observed, a district the Clerk printed no vote total for, and a race with only one major-party candidate. Those flags mark the cleaning that mattered.

38 districts had only one major-party candidate in 2024, so they have no Democratic share of a two-party vote. They cannot be dropped, because a forecast ignoring 9% of the House is not a forecast of the House. Each gets a starting share predicted from the presidential vote inside it, which now tracks the House vote to within 2.5 points. Two further districts arrive with no vote total at all, because the Clerk prints a footnote marker where an unopposed candidate’s count would go. Both are assigned to the party that carried the district for president. That is a rule rather than a lookup.

67 seats start within five points of the line and the other 368 decide nothing. The model itself is three lines. First, the swing: subtract 2024’s 48.65% from the national two-party vote you assume, which the polls make +5.08 points. Second, add that swing to every district’s starting share — a uniform swing, the same shift everywhere — plus two random errors: one drawn once for the country, one drawn separately for each seat. Third, count every district above 50. Doing that once is one imagined election; doing it 20,000 times, with fresh errors each time, is the simulation, and the share of those runs in which the Democrats reach 218 is the probability the forecast reports.

Follow one district through. PA-07 finished 2024 at 49.5% Democratic of the two-party vote, so that is where it starts. Add the swing and it stands at 54.6% before any error is drawn. A national error of two points against the Democrats would pull it back to 52.6%, a district error is then drawn on top, and the seat is counted Democratic in any run where the total ends above 50. This district matters because it is the 218th when all 435 are sorted by starting share: the seat that delivers a majority.

The national error is one number applied to every district, standing for the polls being off or the last three months going unexpectedly. The district error is drawn fresh for each seat, standing for a retirement, a scandal, a local issue. The rest of this brief is what follows from that difference.

Before you read on, write two numbers down. The polls imply a national Democratic share of 53.73%. How wide is the 80% interval around the seat count that implies — how many seats separate the 10th percentile from the 90th? And what probability would you put on the Democrats winning the House? Commit to both before you look.

What it says about democracy

Before any error is added the model is a plain function: give it a national vote share, get back a seat count. That function is the map, and political scientists call it the seats–votes curve.²

Seats the model gives the Democrats at each national vote share, with no error term switched on. The step at each district is one seat changing hands. Hover anywhere on the plot.

Figure 1. Seats against votes on the 2024 map. A step curve fits because seats are won one at a time. Each step is a single district crossing 50%, so the line shows where the map is dense with close seats and where it is empty. The vertical axis is Democratic seats, the horizontal axis the national two-party Democratic share of the House vote, and the dashed line at 218 is a majority.

The Democrats reach 218 seats at 49.2% of the national vote, a margin of R+1.6 — a margin being the gap between the two parties’ shares, here with the Republicans ahead. They can lose that vote by 1.6 points and still hold the chamber, and the seat delivering it is PA-07. The curve runs at about 4.52 seats per point, which is the exchange rate between a poll and a chamber: two points of polling error is roughly 9 seats.

The distribution of Democratic seats over 20,000 simulated elections. Move any dial. Everything below it recomputes.

Figure 2. 20,000 simulated elections. Each bar is a seat count and its height is how often that count came up, with bars at or above 218 a Democratic House. At the polling average the model gives the Democrats a median of 244 seats, an 80% interval of 218 to 275, and a 90% chance of a majority.

Move the first slider and you ask a different question of the same model. Move the other two and you change what the model believes about its own reliability, without touching any data.

Turn off the district error and the interval barely narrows, from 57 seats to 44. Turn off the national error instead and it collapses to 11. Nearly all the uncertainty in a 435-seat forecast comes from one number drawn once. The reason is arithmetic, not politics. The district error is drawn independently 435 times and independent errors cancel; the national error cannot, because there is only one of it. When the polls are two points off, they are off everywhere at once.

That is why professional forecasts rarely report 99%, and why adding district-level detail does less than it looks like it should. Better candidate data shrinks the small term; nothing shrinks the large one except better polls.

What this brief cannot tell you

Which map November will use. Every district here is a 2024 district, and 10 states have enacted new congressional maps since.⁴

The redraws removed 14 competitive seats. The partisan effect roughly cancels; the effect on this model does not, because all its uncertainty lives in exactly those seats.

Which of its inputs is a judgment. The national error is 3.0 points of two-party share, of which 2.21 was measured from the backtest and 2.00 was assumed.

Turning that dial from zero to the default widens the interval by 14 seats and moves the probability by 5 points. The guess matters far more to the range than to the headline, and only the headline is repeated.

Who is running. Every seat’s starting share carries the incumbent who won it, and incumbents retire. Over fifty members have said they will not stand again,⁵ and each of those seats reverts toward its underlying partisanship unseen.

Whether the forecast was any good. In November one thing will happen, and no single outcome shows that 90% was right or wrong. The Senate half is weaker still: its baseline carries no incumbency, and 33 seats cancel their independent errors far less than 435 do.

What you should have learned

• A forecast claims a distribution, not an outcome: 244 seats in the middle, 218 to 275 at the 80% interval, 90% for a majority.
• The map sets the terms. Democrats reach 218 seats at 49.2% of the national vote, so they can lose it by 1.6 points and hold the chamber, and only 67 of 435 districts are close enough to matter.
• Almost all the uncertainty comes from one number drawn once. Switching off the district error narrows the interval from 57 seats to 44; switching off the national error collapses it to 11.
• More data buys a narrower answer, not a different one. Every model here stating a probability lands between 85% and 90%, and one of them is a single sentence.
• The chambers are not independent. Both comes out at 59% rather than the 53% multiplication gives, because a polling miss happens to a country.

Extensions

None of these is answered above, and every one can be answered from the tables Sources hands over.

• generic_ballot.csv holds six aggregators’ numbers in margin and two_party_dem. Which would you feed the model, and what does their disagreement cost at 4.52 seats per point?
• calibration.csv puts err beside nochange_err for eight past elections. Is the model beating the assumption that nothing changes, and in which years does it fail?
• sensitivity.csv varies assumed_poll_error_margin against width and p_house. Find where the answer stops being confident.
• baseline.csv flags districts as imputed, uncontested and no_totals. Count them, and say what share of a national forecast is a guess.
• Stretch. redistricting.csv names the states with new maps. Pick one, find its new lines, and work out how many of curve.csv’s steps near 50% the redraw removes.

Sources

Data

• Clerk of the U.S. House of Representatives, Statistics of the Congressional Election, 2004–2024, for district returns and the presidential vote within each district. https://history.house.gov/Institution/Election-Statistics/Election-Statistics/, every volume from 1920 on. The Clerk’s own address for the series, https://clerk.house.gov/Votes/ElectionStatistics, no longer resolved when this was last checked.
• Office of the Historian, U.S. House of Representatives, Party Divisions of the House of Representatives, 1789 to Present, for the base-rate model. https://history.house.gov/Institution/Party-Divisions/Party-Divisions/
• @unitedstates/congress-legislators, legislators-current.csv (https://unitedstates.github.io/congress-legislators/legislators-current.csv; the project is at https://github.com/unitedstates/congress-legislators), a file of every sitting member kept by volunteers, for every senator’s party and Senate class. That the two independents caucus with the Democrats is not in the file and is applied here.
• Presidential returns by state, 1864–2024, for each state’s lean. https://github.com/jaytimm/PresElectionResults
• Wikipedia, “2026 United States House of Representatives elections” (https://en.wikipedia.org/wiki/2026_United_States_House_of_Representatives_elections), read as the page’s source text, for the six generic-ballot aggregators and the redistricting summary. It is the only place the six appear side by side.
• FiftyPlusOne, “Election forecast: Democrats strongly favored for the House, tilt Senate” (3 August 2026), https://blog.fiftyplusone.news/p/election-forecast-democrats-strongly-favored-house-tilt-senate. Quoted, not recomputed.

The one number that is not from a file

• The assumed polling error, 2.00 points of two-party share (4.0 points of margin). It is a judgment, not a measurement. Bafumi, Erikson and Wlezien (cited above) is where a reader can check how far generic-ballot polls have sat from the eventual vote; the average used here is three months out rather than final, when the miss is larger.

The data itself

All twenty tables are described in The data at the top of this brief. They are plain CSV, they open in a spreadsheet, and they are yours to take further: generic_ballot.csv, baseline.csv, calibration.csv, curve.csv, simulation.csv, sim_default.csv, sensitivity.csv, spread.csv, models.csv, redistricting.csv, senate_seats.csv, senate_calibration.csv, senate_curve.csv, senate_sd_by_year.csv, senate_parse.csv, senate_sim.csv, joint.csv, joint_grid.csv, joint_cloud.csv, joint_compare.csv.

I want to rebuild a public dataset from its original sources, without
any starting files. Please do the following and show your work — and
know one thing going in: part of this source moves every day, so your
copy will be today's page, not the one this book captured.

THE FIRST SOURCE. Wikipedia's page "2026 United States House of
Representatives elections", downloaded as raw wikitext through the
MediaWiki API:
https://en.wikipedia.org/w/api.php?action=parse&page=2026_United_States_House_of_Representatives_elections&prop=wikitext&format=json&formatversion=2
No account or key is needed. Two tables inside it matter: the
comparison of generic-ballot polling aggregators, and the summary of
mid-decade redistricting. Wikipedia is not the authority on either —
it is merely the one place six aggregators who publish no common file
sit in a single table.

THE SECOND SOURCE. The @unitedstates project's roster of sitting
members of Congress:
https://unitedstates.github.io/congress-legislators/legislators-current.csv
One CSV; each row is a current senator or representative, with party,
state and Senate class. No key.

WHAT TO BUILD.
1. The aggregators' national House margins, read out of the wikitext
   table, with their average.
2. From the roster: the Senate seats up in November 2026 — class 2,
   plus any special elections — with each seat's current party.
3. A sanity baseline: the 2024 House result the forecast starts from,
   215 Democratic seats to 220 Republican, with every district given a
   starting vote share.

CHECK YOUR WORK. The copy this book used was captured on 16 August
2026. The aggregator table will have moved since — that drift is the
page being edited, not an error. What should still hold:
- the baseline reproduces the 2024 House: 215 Democratic seats
- all 435 districts carry a starting share
- 537 rows in the roster file, and its senators sum to exactly 100
The roster updates whenever a seat changes hands mid-term; a count a
few rows off means a vacancy or a swearing-in, not a mistake.