State Legislative Forecast

1. Prior. Each candidate's expected vote share combines the district's earlier result (relative to that year's national vote), incumbency, a state baseline and the national environment. Coefficients come from a Bayesian model fitted on 42,338 contested state legislative elections in 44 states, 2000–2022.
2. Information. Official 2026 nominee lists, the latest same-plan district results (MEDSL), current officeholders, and dated generic-ballot polls (VoteHub) for the national environment. Every input is timestamped and listed with its source.
3. Posterior. Joint simulations draw a national environment, national and state shocks and district noise, then add up every chamber. Conditioning keeps only simulations consistent with your scenario; nothing is re-weighted by hand.

No prediction markets, betting odds, news, scandal or subjective candidate-quality inputs. Any future rule-based adjustment has to improve held-out backtests first.

Calibration (district model, held out by state)

Candidate-model backtests (multi-candidate and multi-member seats)

Each fold was fitted only on earlier elections and scored on contests the model never saw. Log loss is for the winner (lower is better); the comparison is a regression fitted on the same features.

2022 was the hardest test: redistricting reset most district histories and the model did not beat the comparison there. The 2024 test was run once, after the model was frozen; most of its gain came from races with third-party candidates. Production rules for missing data use separate empirical rates from Klarner's legislative returns, 2000–2020.