How the projection works

Summary

The model answers one question: if each county votes as it did in 2022, adjusted for how it has moved since and for how the 2026 environment differs, who wins and by how much?

• Anchor. Each county's 2022 margin, taken from certified results.
• Trend. Half of each county's 2024 movement relative to the state is carried forward.
• Environment. One statewide shift, applied most heavily to competitive counties, and not at all in the "2022 pattern" scenario.
• Turnout. Projected from the early-vote pace and spread across counties by their 2022 share of the total vote.

Data sources

The county model

1. Final votes in a county

This asks what share of its 2022 total each county had banked at this point in 2022, and grows the current count by the inverse of that. A county that votes late gets a larger multiplier.

2. County margin

3. Trend adjustment

This matters most in the large metro counties. A county that moved relative to the state between 2022 and 2024 is assumed to have kept moving that way, but only partly, because presidential-year patterns do not fully carry into a midterm.

4. Elasticity

Swing counties move the most and safe counties move the least, which matches how real shifts behave. It also keeps Fulton and DeKalb from being moved by an environment shift that mostly happens at the margin.

5. Votes from margin

The environment shift

The shift s is the one judgment call in the model. It is the statewide change in margin, in points, relative to what the 2022 pattern implies.

Primary electorates are not general electorates, which is why only 15% passes through. The three scenarios on the main page are:

Early-vote pace and turnout

We compare 2026 with 2022 by aligning weekdays, because early voting started on a Tuesday in 2026 and the daily pattern depends on the day of the week. In-person voting ran at 190,405, 174,531 and 163,889 on days one to three, which is 1.41, 1.37 and 1.33 times the same weekdays in 2022. The ratio is fading by about 4 points a day.

Turnout changes totals, not shares. The turnout slider scales every county proportionally, so vote counts change but the margin does not. Differential turnout, such as one party's voters showing up more than the other's, is captured by the environment shift instead.

Governor adjustment

The governor's race uses the same structure with its own 2022 margins (Kemp +7.5 statewide) and Libertarian shares. The early-vote composition is adjusted by each county's gap between advance and final results in the Senate race, because early voters in 2022 were not a representative slice of the final electorate.

The governor's race is closer than the Senate in the polls (R+1 to R+2, with Big Data Poll showing Bottoms ahead by 1 after leaners), and the model's base case is Jackson +0.3, with neither candidate above 50%, which would send the race to a runoff. Treat it as a toss-up.

Reconciliation checks

The 2,346-ballot gap in the absentee line is a timing and reporting difference between the absentee file and the county results, not a modeling error.

Sensitivities

Each row changes one assumption and holds the rest at base.

The pass-through is the assumption that moves the result most. Going from 10% to 20% moves the Senate margin by about 3.4 points and decides which way the governor's race leans. The governor's race crosses zero at about a 16% pass-through.

Prediction markets

The main page reads the Kalshi winner markets for the Senate and governor's races. Each price is the cost in cents of a contract that pays $1 if that candidate wins. We normalize the two candidates' prices so they add to 100 and use a 7-day average to smooth single trades.

The comparison table on the main page uses that one conversion for all three views, so margins and chances are on the same scale. Changing σ moves the chances but not the order of the views.

Limits: the markets are thin, so a single trade can move a daily price. Prices reflect traders' views, which include the same polls and early-vote data we use, and they can be wrong. Market data is a point-in-time download and is refreshed with each update.

Limits

• A scenario is not a forecast of record. The model shows what follows from stated assumptions. It does not attach probabilities.
• Early-vote party is not recorded. Georgia has no party registration, so we cannot see who has voted, only where.
• The pass-through is a judgment. The primary swing is real, but how much of it holds in November is uncertain.
• Polls differ in method. The average is a simple mean of the three headline margins (first ballot, undecideds as reported), not adjusted for sample size, population or house effects. InsiderAdvantage had not posted Senate toplines when we built this, so its Senate margin comes from the release headline.
• Runoff. Georgia requires 50% plus one for the general election. If neither candidate clears it, a runoff follows about four weeks later.
• Data revisions. Early-vote files can be corrected after publication. Official results come only from the Secretary of State and county election offices.

County data

The inputs for all 159 counties. Projected votes use the base scenario and the central turnout assumption.