For each of the 27 districts: who will be elected senator, either by an absolute majority in round 1 or by winning the runoff. Each district gets a probability distribution over the main candidates plus “others”, summing to 100 %. Resolution follows the official ČSÚ result (volby.cz).
Stages and publishing
Blind: a forecast made without seeing betting odds. Published first.
Final: the blind forecast combined with the market (odds with the bookmaker's margin removed). Published separately and later; the blind number does not change.
Updates: new versions when new information arrives (e.g. after round 1). Every version is a new record; old ones are never overwritten.
Every day an immutable snapshot goes to the archive and a commit goes to git. Corrections are published as new versions with an explanation, never as silent edits.
What gets scored
Primary scored snapshot: the last version published before voting opens, i.e. before 9 Oct 2026 14:00 CEST. Blind and final are scored separately.
Post-round-1 updates are scored separately as secondary. Forecasting gets much easier after round 1, so they don't count toward the main score.
Metrics: averaged over all 27 districts:
multi-outcome Brier score, Σ(pi − oi)²; lower is better;
log score of the winner, ln p(winner); closer to 0 is better;
number of districts where our favourite won.
Benchmarks:
de-vigged betting odds, where available;
a naive rule, “the incumbent wins, otherwise the candidate with the broadest coalition” (Brier and favourite count only);
a uniform split across the main candidates.
Important caveat: correlation
The 27 districts are not independent trials. The national mood, for example discontent with the government, moves all districts at once. In most districts the favourite is an opposition candidate, so the outcome is largely one big correlated bet. We will say so in the evaluation.
How the forecasts are made
Question normalization.
Research: candidates, district history, 2025 Chamber results, local context.
Blind forecast: a model of round-1 shares and runoff pairings, with win probabilities conditional on the opponent and on who led round 1.
A grounding check that every claim rests on the research.
Red-team.
Revision and aggregation.
Shared base rates:
The round-1 leader wins the runoff about 70 % of the time (ČSÚ).
ANO candidates won only about 9 of 44 runoffs in 2020–2024, and about 1 of 8 in 2020, when ANO was in government.
Five series (from 5 Oct)
AI: the pure AI agent, with no model and no odds. It changes only on new facts (news, endorsements, scandals, polls). The “AI 2 Oct” column shows the starting point.
Statistical model (see model description). It changes only when its inputs change, e.g. after round 1.
AI + model: the AI agent starting from the model and deviating only for facts the model cannot see. Every deviation is logged. No odds.
Betting: Fortuna and Tipsport seat-winner odds with the margin removed (power method). Where Tipsport only offers a yes/no on one candidate, we average it with Fortuna and split the rest in Fortuna's ratios. Odds are read only after the blind series are locked.
Blend: a mechanical log-odds combination of “AI + model” and “Betting”, ρ = 0.55 (wblind = 0.31). No judgment, just a tool.
AI, Model and AI + model are blind. Their snapshots are made before any odds are read and are locked in git before the odds are added. Any later AI update made after seeing odds will be labelled post-market. All five series are scored.
What we don't publish
Betting recommendations or stakes. This is not a tipping service. We don't publish raw working files either.
Known limitations of the first run (2 Oct)
The web-search budget ran out mid-run. For most districts, research relied on directly fetched pages (Wikipedia, programydovoleb.cz, senat.cz). Fresh local news and often 2025 district-level results are missing. District pages flag this as research_limitation.
Because of a concurrency limit, the red-team in 4 districts (12, 21, 24, 39) did not run independently. We therefore ran independent red-teams separately before publication.
Several forecasts were revised after the red-team before publication. The published state is the starting point; from here on only new versions count.
The blind forecast does not use opinion polls systematically. District-level polls are rare, and with the search budget exhausted most districts could not search for them properly. “No poll” therefore means “none found” rather than “none exists”. National polls (STEM, Median, NMS, Kantar) served only as context, not as a model input. Plan: by Monday 5 Oct (a poll-publication blackout starts on 6 Oct), search specifically for district polls and apply the national swing since the 2025 election uniformly to all districts as a new update.
Plzeň (district 9) had already gone through the market step. We publish only its blind number here; its final comes with the others.
Timestamps
Each daily state is a git commit plus an archive.org snapshot, so it can be shown that a forecast existed before the result.