Skip to methodology
Donate

How the NavCast works

The NavCast estimates where an election is heading as votes arrive. It combines a frozen starting forecast, historical precinct patterns, reported votes and estimates of the vote still outstanding for Governor, State House and State Senate races.

The current release candidate is an experimental forecast evaluated against official historical final tallies. Its interval coverage and errors are measured below. Limited election samples and simulated reporting do not establish a guarantee for the next election.

Methodology reviewed October 7, 2026. This page describes the local Georgia release candidate. Live source qualification and deployment are separate from this review.

Where the starting projections come from

The Georgia baseline is frozen from the October 1, 2026 State Navigate forecasts. House and Senate precinct major-party shares blend the 2024 legislative result (69.2%) and the 2024 presidential result (30.8%). These are the relative weights of those two historical inputs, not the weights of the full forecasting model. Historical legislative shares supply the local shape and other-party share; total turnout is set at county level using midterm history. Each district’s precinct estimates are then adjusted to reproduce its 2026 forecast margin.

For Governor, the county and statewide forecasts supply the starting targets. The current precinct shape uses 2024 presidential results, with voting-method composition from the 2022 Governor county results. The available 2022 Governor export did not supply precinct rows, so this is an explicit proxy.

County turnout

Georgia’s turnout baseline uses county-level midterm patterns, rather than carrying presidential-year legislative vote totals into 2026. It compares 2018 Governor votes with 2016 presidential votes, and 2022 Governor votes with 2020 presidential votes. Each county’s 2024 presidential vote total supplies the current size anchor. The default combines the 2018 pattern at 35% and the 2022 pattern at 65%, with small-county ratios pulled toward the statewide pattern.

Governor is the shared turnout anchor. House and Senate totals apply county vote-retention rates from the complete 2022 legislative returns, so both remain below Governor. The starting totals are 4,467,858 for Governor, 4,113,177 for House and 4,079,298 for Senate. County totals are allocated across the existing precinct catalogue; the 2026 race-margin forecasts are retained. Official Georgia historical returns.

The simulator can use the blended baseline, a 2018 or 2022 county pattern, or a lower-turnout 2014 stress case. These are explicit scenario assumptions, not calibrated confidence limits. The generic ballot does not automatically select a turnout pattern. County partisan groups describe where votes are cast, not the turnout of individual party supporters.

State Navigate precinct geometry, county identifiers and historical election exports help connect the baseline to the results feed. A historical precinct split across districts is treated as a split proxy; its votes are not copied into multiple races as additional ballots. A new baseline is a new frozen model version, rather than an undocumented change to an election-night projection.

How votes arrive

The model keeps those voting methods separate whenever the source supports them. Reported counts remain observations; historical allocations and estimates remain projections. Missing values remain unknown, and a reported zero remains zero.

A completed voting method fixes its known counts while other methods can still have votes outstanding. A partially reported precinct has less influence than a completed method. Partial totals without a usable method breakdown do not supply partisan shift evidence, because the mixture of early and Election Day votes can change as counting proceeds.

Turnout updates are learned primarily from completed methods. A small first batch does not establish low final turnout, and reaching the starting turnout expectation does not establish completion. Partial reporting can exceed that expectation while retaining an outstanding-vote tail and extra uncertainty. Pre-count participation may update expected method volume after early voting closes, using official dated counts. Cast ballots do not become counted candidate votes. Current registration bounds apply once per office and county, across its district portions, and never erase known votes.

Learning from comparable returns

The selected updater learns an office-wide shift separately for early in-person, mail, Election Day and provisional votes. Vote totals determine the weights, so large counties contribute according to their ballots. Prior population coverage and the effective number of reporting counties restrain evidence concentrated in a few places.

Each race and county can depart from that shared movement. A matched precinct also uses its own method-specific return, with weaker weight for small counts. Borrowed shifts are capped; a sufficiently large observed change in the precinct itself can exceed that borrowing cap. Only votes still outstanding receive the adjustment.

Governor results do not transfer their partisan margin directly into legislative races. House and Senate districts learn from their own office and returns. Completed turnout can inform another office in the same physical precinct. The selected conservative updater does not copy partisan movement between voting methods.

Demographic features in the starting data are district proxies rather than precinct observations. The conservative partisan update uses office, race, county and method evidence; these proxies do not create new demographic measurements.

A worked example

Suppose early votes were expected at D+10 but arrive at D+16: a six-point Democratic margin shift. This example has two equally sized counties with 10,000 expected early votes in total, and only one county reporting. The unreported county starts at D+2. The calculation shows the shared component after population coverage and geographic concentration reduce the donor’s influence. It omits local adjustments and uncertainty.

QuantityValue
Reporting precinct’s expected early marginD+10.0
Reporting precinct’s observed early marginD+16.0
Neighbor’s starting early marginD+2.0
Eligible evidence vote mass1,000
With 1,000 votes and a completed method, the shared component moves the remaining early-vote margin from D+2.0 to about D+4.8.

Only votes still outstanding receive this adjustment. Reported votes stay fixed. The six-point signal is a margin change, equivalent to a three-point vote-share change in a two-party race.

What happens when precincts change

The current local implementation accepts partial precinct crosswalks. A match must use a verified source identifier within the same race and county. Similar names alone do not establish a match. New precincts, renamed identifiers and county mismatches are retained as unmatched observations rather than stopping the model.

Their reported counts contribute to county and race vote bounds, but they supply no partisan residual, turnout donor or independent statistical observation. Unmatched baseline units continue from their prior estimates, updated by comparable matched returns, while retaining mapping uncertainty. The results table can show an unfamiliar precinct’s actual votes without inventing a map boundary or a precinct projection for it.

Parent and child totals are reconciled without adding both together. A lagging county total cannot discard newer precinct votes. Counts that cannot be allocated to a matched precinct consume expected remaining turnout before being added with their actual party totals. That bookkeeping allocation stays separate from precinct observations and shift evidence.

What the uncertainty means

NavCast displays the central 90% prediction interval for final margins and chamber seat counts, using the 5th and 95th percentiles of the model's simulated outcomes. The Governor forecast's modeled uncertainty supplies the starting statewide distribution, which updates as returns arrive. The model also retains 50%, 80% and 95% intervals for evaluation. A 90% interval does not mean a party has a 90% chance of winning. Whether intervals attain their stated coverage must be measured against completed elections. Prediction-interval interpretation.

Uncertainty covers the starting forecast, county and voting-method shifts, remaining turnout, mapping and allocation weaknesses, model discrepancy, and possible unofficial tally corrections. Effective precinct and county clusters determine the information gained; thousands of ballots from one reporting cluster are not thousands of independent observations about the rest of the state.

The model uses 10,000 correlated, heavy-tailed simulations. It retains forecast correlations and shared turnout factors, so the chamber forecast does not assume that every close district moves independently. Separate method uncertainty prevents a large early-vote batch from removing all Election Day uncertainty. Completion without explicit certification evidence retains a correction allowance.

Governor county rows and House and Senate district rows show 90% prediction intervals by default. Governor county party-vote draws sum to the statewide draws. Legislative district draws determine chamber seat draws. Bounds account for remaining turnout, voting methods, matching uncertainty and unofficial corrections. Marginal county interval endpoints cannot be added to obtain a statewide interval. A county covering an entire race shares that race's bounds.

The candidate uses residual scales fitted on earlier elections, with a separate late-method profile where earlier evidence supports it. A voting method gaining a much larger share of voters retains an additional correlated reserve. The candidate preserves every starting race forecast and has a separate baseline identity from the earlier prototype.

Race and chamber probabilities

D/R race probabilities currently describe which major party finishes ahead of the other. Outright-majority and runoff estimates are separate. They do not resolve an individual minor-party winner or the winner of a later runoff. Georgia requires a majority to avoid a runoff, and a first-place finish can fall short of that threshold. Georgia runoff guidance.

Chamber odds are conditional on the simulated first-place winners. The display uses five non-overlapping categories: Democratic supermajority, Democratic majority below the supermajority threshold, a tied chamber, Republican majority below the supermajority threshold, and Republican supermajority. Expected Democratic seats can be fractional because they average many simulated outcomes.

The NavCast does not issue automatic race calls. Its current exports leave calls unset. A high modeled probability, a vote lead, or a precinct reporting flag does not become a called race.

What the historical evaluation shows

Governor residual scales use 2018; legislative scales use 2014, 2016 and 2018. Evaluation uses 2022 Governor results and 2020 and 2024 legislative results. The Governor comparison uses archived FiveThirtyEight starting margins. Legislative priors are historical benchmarks, including past Governor margins on the current district geography for the 2024 diagnostic.

The table excludes races with at most 1% of actual votes left and gives each election equal weight for coverage. Thousands of repeated county or district rows are correlated observations, not thousands of independent elections. Governor has one evaluation election and each chamber has two.

Scope80% coverage95% coverageMargin error
Governor counties89.6%97.8%2.08 points
House districts82.2%96.3%2.92 points
Senate districts87.0%97.6%2.65 points

Reporting sequences include early-first, Election Day first, late mail, delayed large counties, geographically concentrated returns and missing methods. Reporting order is simulated. Later outcomes were consulted during diagnosis, so this is retrospective chronological evaluation rather than an untouched external test. The late 2020 stages remain the weakest part of the evidence.

Implementation checks separately exercise counts, corrections, completion, matching, coherent draws and publication recovery. Passing software tests alone does not establish statistical calibration. The dated evidence includes interval widths, interval scores, probability scores, election counts and source fingerprints. Past-to-future evaluation.

Updates and the time chart

The worker is designed to check the official source every minute during an enabled election window. Changed, validated input triggers a new projection. Rechecking unchanged counts updates collection health without manufacturing a new model point. A failed download, invalid candidate catalogue, wrong election or incomplete upload retains the last complete release. Current results and projections load first; county or district history and precinct detail load when requested. Demonstrations are labeled and cannot enter the live feed.

The page’s time chart follows projected and called seat counts, chamber outcome probabilities, and estimated vote counted through accepted updates. Estimated vote counted uses projected final turnout; it is not the percentage of precincts reporting. Corrections can move counts and estimates downward. Demo history records saved simulated updates, rather than reconstructed live election-night history.

Source responses, retrieval times, input versions and release hashes are retained for inspection. Test data, historical replay and live data have separate identities. The Georgia election pages check for accepted releases every 30 seconds; the source itself determines when new vote totals become available.

How the model should be used

The candidate is suitable for a public experimental forecast with its evidence and limitations visible. Its safeguards address uneven reporting, voting methods, geographic concentration, unknown precincts and corrections. The measured intervals do not promise 95% coverage in a new election, and precise win probabilities are not independently certified.

Before live use, the election identifier, candidate fields and known precinct matches must be qualified against the actual source. The release candidate remains local until it is deployed. Georgia is the first implemented adapter; listing 22 A-tier states does not establish working or qualified live feeds for all of them.

Additional independent elections and actual archived reporting timelines would strengthen the evidence. The current comparison includes the earlier updater, a vote-weighted uniform method shift and a restrained Governor county trend. The conservative updater is selected for this candidate; the trend is retained as research rather than changing the frozen starting county shape. Evaluation with past-to-future validation.

Changes to model assumptions, mappings or source identity should be documented and versioned. Official reported counts remain the authority for what has been counted; the NavCast estimates what may remain.

Sources and further reading