Prediction-market calibration by probability band and horizon
This first Phase 4 research report documents a frozen calibration data run: probability bands, horizons, sample definition, exclusions, and a versioned query reference. Narrative claims stay limited to what the frozen package supports.
Frozen data run
- Date range
- JUN 1, 2026 – AUG 31, 2026
- Sample
- Settled SEO catalog market-event pages with timestamped primary-outcome observations inside the pre-resolution lookback, frozen at data-run time.
- Exclusions
- Suppressed/test/canceled-without-result rows; unknown outcome semantics; short recurring props; incompatible cross-venue merges.
- Query / code ref
- insydr.research.calibration_by_band.v1 (frozen SQL/package id)
- Frozen at
- SEP 9, 2026
Aggregate CSV download will attach when the freeze package is published; methodology and sample rules above already define the reproducible frame.
Why this report exists
Original research is Insydr's strongest citation asset. This note automates the data package and sample rules, not an unsupported narrative. It explains how we will measure calibration of exchange-implied probabilities against later realized outcomes, by probability band and by time-to-resolution horizon.
Methodology
We map each included market-event page to a primary outcome probability observed at a fixed lookback before resolution (or before a defined horizon cutoff). Probabilities are exchange-reported prices from Insydr snapshots (decimal 0–1), never invented blends. Outcomes are taken only from provider-confirmed settlement on archived pages.
Bands are fixed: 0–10%, 10–20%, …, 90–100% implied probability at the observation point. Horizons are fixed buckets (for example under 7 days, 7–30 days, 30–90 days, over 90 days to resolution). Within each band×horizon cell we count observations and realized yes rates when the sample is large enough to report.
Sample definition
Included rows are public catalog market-event pages that were indexable or archived-indexable at freeze time, had a settled lifecycle with a confirmed result, and had a usable primary-outcome probability observation inside the lookback window. Multi-outcome events use the primary tracked component defined by the snapshot builder.
Exclusions
Excluded: test or suppressed pages; canceled markets without a usable result; short recurring props tagged out of the SEO catalog; observations missing timestamps; contracts whose outcome semantics are unknown; and any row that fails content-safety or completeness gates at freeze time. Cross-venue averages are never computed across incompatible definitions.
Limits
Calibration is descriptive, not advice. Thin cells are omitted or marked unstable. Past band performance does not imply future accuracy. This HTML page is the reviewed publication surface; the aggregate CSV (when linked) is the downloadable package for the same freeze.
Corrections
Material corrections update this page in place, bump `updatedAt`, and add a short correction note when a prior error was significant.
Sources
Corrections and methodology questions: Contact.