Observed setup
The audit observed four horizons: 6h, 12h, 24h, and 48h, generated on a sixth-hour cadence.
Public forecast documentation
Forecast methodology and metric definitions
These definitions describe how the audited statistics should be interpreted. They do not claim predictive superiority or identify an unverified model implementation.
Historical audit snapshot created 2026-09-10 — not a live/current forecast
Overlapping samples
Because longer horizons exceed the six-hour issue cadence, their evaluation windows overlap. Adjacent errors are therefore not independent observations.
Direction Brier score
For n resolved forecasts: (1/n) Σ(pᵢ − yᵢ)². Here pᵢ is probability of up and yᵢ is 1 for up, otherwise 0. A constant p = 0.5 scores 0.25.
Normalized absolute error
For each resolved forecast: |predicted price − observed price| / |observed price|. An aggregate must state whether and how these ratios are averaged. It is not raw-price MAE.
80% confidence band
A lower and upper forecast interval with nominal 80% coverage. The label describes a long-run calibration target, not certainty for an individual outcome.
Resolved forecast
A forecast whose target timestamp has passed and whose outcome price, scoring source, and scoring rule have been recorded.
Data provenance and recommended public export
The statistics shown here come only from the forecast audit snapshot created 2026-09-10. Current values remain on the external live product. The local scheduled-worker forecast table is a distinct implementation and is not represented as that external production feed.
Recommended source: a versioned, downloadable JSON and CSV dataset at the live product, with immutable issued and resolved rows, UTC timestamps, source venue, sample count n, scoring formulas, model version, and a content hash. The Dataset JSON-LD on this page preserves the audit date rather than implying freshness.