Model v1.2.0
How the forecast works
Public data in, 40,000 simulated elections out.
1Collect
Polls, expert ratings, candidates, money and the national mood.
Poll average: Cooper +9
Cook: Lean D (flip)2Blend
Fresh polls count most. Without polls, the basics decide.
Basics: Cooper +8.2
Blended: Cooper +8.93Simulate
Play out the election 40,000 times, with realistic poll misses.
9 in 10 runs land between Whatley +0.4 and Cooper +18.14Count
Our number is the share of those runs a candidate wins.
Cooper wins94%
Example: North Carolina Senate › · RCP average, Sep 30 2026
What goes in
Five kinds of public data, each one cited.
Not used: prediction markets, social media or early voting.
Polls and the basics
Every race starts from the basics: the expert rating turned into points, a small boost for incumbents, a nudge toward the side raising more money, and part of the national mood.
A recent poll average then takes over. Its weight halves every 30 days:
Most House races have no public polls. There, the forecast is the basics alone, with a wider range.
The formula
- Rating to points: Solid 20, Likely 5.5, Lean 3.5, Tilt 1.5, Toss-up 0. The Solid figure is an interim value, set from how those states vote; we haven't yet tested it on past elections.
- Incumbent: 2 points toward their side.
- Money: one point per tenfold fundraising gap, capped at 1.5.
- National mood: half the generic-ballot margin (D+8.4 today), because the ratings already reflect some of it.
- Forecast = weight × poll average + (1 − weight) × basics, where weight = ½ ^ (age in days ÷ 30).
The half-life, incumbency boost, money cap and the halving are judgment calls, not fitted from data.
When polls miss, they miss together
In 2016 and 2020, polls missed in the same direction almost everywhere. So each simulated election draws one national miss that moves every race at once, plus a smaller miss for each race.
One national miss moves every race the same way.
That is why our Senate number is 73% and not higher: one bad polling year can swing several close races together.
How big the misses are
- National miss: typically 2.5 points, shared by every race.
- Total miss for a polled race: it shrinks as the election nears and as more polls come in (polls ending in the last 21 days). The race's own part is whatever the national part leaves.
- Races with no polls: 7 points of their own, on top of the national miss.
| Total miss, in points | 1 poll | 2 to 4 | 5 or more |
|---|---|---|---|
| Last 3 weeks | 8.2 | 6.5 | 4.9 |
| 3 to 6 weeks out | 9.7 | 7.1 | 5.6 |
| 6 to 9 weeks out | 9.8 | 7.4 | 5.8 |
Fitted from FiveThirtyEight's archive of Senate, governor and House polls, 1998–2022. Tested on each cycle from 2006 using only earlier cycles, it beat the single fixed size we used before. The national part came out at 2.1 points and was rounded up. The figure for unpolled races is a judgment call.
40,000 elections
We play out the whole election 40,000 times. A candidate's chance is the share of runs they win. Control of each chamber is counted from the same runs, so the numbers always agree.
Runs where Democrats reach 51 seatsOut of 40,000
73%
The runs use a fixed seed, so the same inputs always give the same numbers.
Run details
- Model v1.2.0, run Oct 9, 2026, seed 20261002.
- 470 races: 35 Senate, 435 House.
- Senate: 34 seats not up this year count for Democrats and 31 for Republicans. Democrats need 51, since the vice president breaks ties. House: 218.
- Re-running with a different seed moves race chances by 0.16 points on average, 1 at most.
- A race's 90% range is its forecast margin plus or minus 1.645 times its total miss size.
How we check ourselves
Before launch, we tested the uncertainty math on the 2022 Senate races, using polls from the final three weeks. The side it favored won 26 of 28. That's a test of the method, not a track record.
Our real record starts with 2026: the full forecast has never faced a real election. Every forecast is saved when we publish it and graded once the votes are counted.
Track record ›Scores and known limits
2022 check: a Brier score of 0.052, where always saying 50% scores 0.25 (lower is better). Both misses, Georgia and Nevada, had polls that underestimated Democrats. 2022 also helped fit the miss table, so the fairer test is the one above: each cycle from 2006, scored using only earlier cycles.
- One poll average per race; no pollster ratings yet.
- Misses are linked nationally, not yet by region.
- The national mood moves every race by the same amount.
- Ratings already reflect money and incumbency, so those terms are kept small.
- Nebraska's independent, Dan Osborn, counts on the Democratic side for control.
Sports and world events
The same four steps run our other numbers, with their own inputs: team ratings and the injury report for games, and the market price blended with a base rate for world events.
Built on public work by FiveThirtyEight, The Economist and independent forecasters.