Our contest entry
The HKS contest scores the House only: the number of Democratic seats, the chance of a Democratic majority, and the chance the Democrat wins five districts. These are the numbers our model gives today.
(c) Probability the Democrat wins each district
Why we enter the median for (a)
How the tiebreaker is scored
If teams tie on (a), the six probabilities from (b) and (c) decide. They are scored with the Brier score: the average of (probability − outcome)², where the outcome is 1 if the Democrat wins and 0 if not. Lower is better.
The Brier score rewards honest probabilities. Pushing a number toward 0 or 1 only pays off if the race really is that certain, so we enter the model's probabilities as they are.
Try your own weights
Re-runs the House model in your browser, 10,000 simulations| Your settings | Our final model |
|---|
Browser runs use 10,000 simulations; the official entry above uses 20,000, so the two can differ by a seat or a point of probability.
From the course model to ours
One change per step, same random draws throughoutOur assumptions
Every number the model needs that the data does not give us: its value, where it enters, how we got it| Assumption | Value | Where it enters the model | How we derived it |
|---|
How this forecast is made
Full method and sources-
Three kinds of evidence for every race
-
Weighted by what each source knows
-
20,000 simulated elections
-
Chances, seat ranges, deciding races
The chance of control is the share of simulations a party wins. The seat ranges and the races that decide control come from the same simulations.
Who wins each chamber
House
435 seats · 218 for a majorityDemocrats–
Republicans–
Median Dem seats
–
80% interval
–
95% interval
–
Senate
35 seats up · Democrats need 51, Republicans 50Democrats–
Republicans–
Median Dem seats
–
80% interval
–
95% interval
–
Why 20,000 simulations, and how stable are these numbers?
What do the intervals mean?
Each simulation ends with a seat count. The 80% interval holds the middle 80% of those counts: in 8 of 10 simulated elections, Democrats end up inside it. The 95% interval holds the middle 95%. The median is the count in the middle, with half the simulations above and half below.
The Senate counts include the 34 Democratic and independent seats that are not on the ballot this year. Democrats need 51 seats because Vice President Vance breaks a 50–50 tie for Republicans.
Control of Congress
Which races decide control
How often each race is the tipping point, races ordered from most Democratic to most RepublicanSenate
House
What is a tipping-point race?
How many seats
Share of simulations ending with each number of Democratic seatsHouse: Democratic seats
Senate: Democratic seats after the election
Races to watch and seat map
Select a race or a state to list it belowSenate
House
What do safe, likely, lean and toss-up mean?
The labels translate our chance of winning into words. They are the same for both parties:
Toss-up · 45% to 55%
Lean · 55% to 65%
Likely · 65% to 95%
Safe · 95% or more
These bands are narrower than most forecasters use. Split Ticket, for example, calls a race a toss-up until one side reaches 60%, lean up to 80% and likely up to 95%. Narrower bands give more races a direction. Even so, the favourite in a Lean race still loses 35% to 45% of the time.
The bands are fixed cutoffs on the chance of winning. They have nothing to do with how uncertain the polls are; that uncertainty only decides which band a race lands in.
Senate: 35 races
House: 435 districts, one square each
Where polls and markets disagree
Races where the poll average and the market price differ by 8 points or moreNational environment
The national House margin the fundamentals start from, and the signals beside itIf the polls miss
Results grouped by the size and direction of the shared national error| Scenario | Chance | Dem House | Dem Senate | Median House D | Median Senate D |
|---|
Every race
| Race | Held by | Chance Dem wins | Polls | Markets | Fundamentals | Dem margin, 80% range | Rating |
|---|
Other forecasts
For comparison only. Other models are not inputs to this forecast.| Source | Dem House | Dem Senate | As of |
|---|
How the forecast works
We forecast each of the 470 races on the ballot, 435 House seats and 35 Senate seats, and then play the election out 20,000 times. Each race draws on three kinds of evidence: polls, prediction markets and fundamentals. No kind of evidence is favoured in advance. Each counts according to how much it knows about that particular race. Polling misses and national swings hit many races at once, so the simulation moves races together instead of treating them as independent coin flips.
- Polls. Newer polls, larger samples and likely-voter samples count more. Polls paid for by a party or campaign count half. A pollster's older polls in the same race count less. Polls of hypothetical matchups with candidates who lost their primary are dropped.
- Prediction markets. Prices from both venues are averaged by trading volume. A market counts like a poll average: $20,000 traded is worth one poll, $1 million or more is worth ten, and never more than ten.
- Fundamentals. For the House we use Decision Desk HQ's district probabilities of 22 October, the course's starting point. Our own version, the 2024 presidential result on this year's lines plus a uniform national swing, missed badly in places (see the bridge). For the Senate: the 2024 presidential result in the state compared with the nation, plus the national environment of (polling average and Kalshi's market on the national House vote), plus 3 points for an incumbent.
- Combining. Each input becomes an expected margin with its own uncertainty. The three are averaged with weights inversely proportional to their variance. A race with many fresh polls leans on polls; a district with a liquid market but no polls leans on the market. Because the inputs overlap (markets read polls, for example), no race is treated as more certain than ±4 points. Hover over any race to see its weights.
- Simulation. In each of 20,000 simulated elections, every race moves with a shared national error of ±3.6 points (the generic ballot's average miss in past midterms), drawn with fat tails so that large misses stay possible. Each race also gets regional, state and race-specific errors. We then count seats. A win by an independent, such as Nebraska's Dan Osborn, is not counted as a Democratic seat.
- What we leave out. Apart from DDHQ as the House fundamentals, other forecasters' models and expert ratings (Cook, Sabato, Inside Elections) are not inputs. We read them to learn and show them under "Other forecasts". DDHQ itself already contains some market information, so markets count slightly more than the weights show.
Process
- 26 OctThe tournament opened. We researched data sources, the 2025–26 redistricting, and what is known about midterm polling errors.
- 26 OctGround rules set: no source type favoured in advance; neighbour and expectation polls kept as a national signal; large poll–market gaps reviewed by hand; other models used only for comparison.
- 26 OctData pipeline built (markets, polls, gap report, fundamentals) and the first model run. Two corrections along the way: markets were initially weighted above polls, and the national error was too small.
- 27 OctPage redesigned; robustness check with four extra simulation runs; explanations added.
- 27 OctContest rules and assignment arrived. Missing markets found (NC-11 and ten others). Bridge built from the course's Part 3 model to ours. Backtest on the midterms 1998–2022 and the 2024 Senate markets.
- 27 OctFinal model v2 chosen: error sizes from the backtest, DDHQ's probabilities as the House fundamentals, all three sources weighted by information.
Every step and choice is logged with its reasoning in the project's process document (PROCESS.md).