Inventory
What the record holds
Records per domain measured 2026-10-29. Older miss and episode counts remain separately dated 2 Oct 2026.
Records per domain
- Economics1,979,271
- Health567,189
- Market traders457,874
- Crypto255,807
- Ensemble weather225,235
- Single-run weather205,746
- Prediction markets170,096
- Central-bank surveys156,912
- Hurricanes152,707
- Market prices119,744
- Space weather103,623
- Geopolitics & world events56,013
- Energy137
Bar length on a log scale so small sources stay visible; numbers are exact.
Frozen miss counts · 2 Oct 2026
Confident and still wrong
Cases where the forecaster's own range was among the narrowest of at least three peers on the same outcome, and still missed.
- Weather35,970
- COVID7,595
- ECB survey2,494
- Flu1,322
- RSV464
Frozen health episodes · 2 Oct 2026
Whole training episodes, health
One episode is one outcome as seen at the decision time. Weather, ECB, ForecastBench, crypto, markets and sea have episode feeds too.
- COVID8,554
- Flu5,070
- RSV4,507
Sources and access
Inventory · 2026-10-29Coverage counts use different units; they are not a ranking of training quality. Open a source for its unit and recorded coverage.
Geopolitics & world eventsForecastBench56,013 Details +
answered question-horizons · Jul 2024 – Sep 2026. Use row-level admission and outcome checks before training.
Open evidenceSingle-run weather7 model centres205,746 Details +
cases · Jun – Oct 2026. Use row-level admission and outcome checks before training.
Open evidenceEnsemble weatherOpen-Meteo ensembles225,235 Details +
cases · Sep – Oct 2026. Use row-level admission and outcome checks before training.
Open evidencePrediction marketsManifold, Polymarket, Numinous170,096 Details +
resolved questions with prices (97,788 Manifold, 71,195 Polymarket, 1,113 Numinous) · Aug – Oct 2026. Use row-level admission and outcome checks before training.
Open evidenceMarket tradersPolymarket wallets457,874 Details +
trades · 2026. Use row-level admission and outcome checks before training.
Open evidenceCryptoKalshi, Deribit, reference models255,807 Details +
forecasts (mostly replays) + 51,459 venue quotes · Jul – Oct 2026. Use row-level admission and outcome checks before training.
Open evidenceCentral-bank surveysECB SPF156,912 Details +
probability distributions · 42 rounds. Use row-level admission and outcome checks before training.
Open evidenceEconomicsPhiladelphia Fed SPF1,979,271 Details +
survey responses (2,250 first-release outcomes, RGDP/PGDP only) · 1968 – 2026. Use row-level admission and outcome checks before training.
Open evidenceEnergyNESO demand vs Elexon outturn137 Details +
resolved half-hours (1,449 admitted forecasts at release) · Oct 2026. Use row-level admission and outcome checks before training.
Open evidenceHealthCDC FluSight, COVID-19 and RSV hubs567,189 Details +
distinct dated outcomes (12,433 flu, 536,748 COVID, 18,008 RSV) · 2020 – 2023. Use row-level admission and outcome checks before training.
Open evidenceSpace weatherKp observations103,623 Details +
observations · Sep – Oct 2026. Use row-level admission and outcome checks before training.
Open evidenceHurricanesNOAA NHC ATCF forecast decks + best track152,707 Details +
forecast rounds (one per method and cycle) + 1,711 best-track rounds · 2024–2025 seasons. Collected coverage; not yet a scored training total.
Open evidenceMarket pricesPolymarket price history119,744 Details +
pre-deadline price points on 19,019 temperature markets · 2026. Use row-level admission and outcome checks before training.
Open evidenceBar lengths use a log scale. Source-specific daily volumes are not yet measured; these bars are dated totals, not trends.
Partner feed freshness · 24 hours
Captured notflat.world health checks, not Offdiagonal collector totals. Each line shows minutes since the upstream update; rising means older data, not more records.
Reading measured histories…
Missing ages and skipped hourly checks break the line. Daily record-volume history is not measured by this feed.
Access
Request data access
Catalogue, method and the example report are public. Full datasets are released against a scoped credential, so each consumer's reach is stated and revocable. A person reads every request.
Request
Tell us what you need
What is public without a credential?
What is open
Without any credential
- The published benchmark: population, admission rule, coverage, abstention and uncertainty.
- The contribution funnel: what each source contributes and where records are lost.
- Worked examples chosen to be representative, not flattering.