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
What this methodology page establishes
This page is the authoritative guide to interpreting the forecast process and its evaluation vocabulary. It separates how forecasts are described and scored from whether the available evidence shows useful performance.
The main SiCierto site does not have a configured, authenticated upstream feed for the independently deployed forecast product. It does not substitute the local scheduled-worker database or label the dated audit values as live.
Setup observed in the audit
The September 10, 2026 snapshot observed four forecast horizon labels — 6h, 12h, 24h, and 48h — issued on a six-hour cadence. This is audit context, not a claim about a current production schedule.
- Forecast horizons
- 6h, 12h, 24h, and 48h labels were observed.
- Issue cadence
- Six hours in the audit observation.
- Evaluation period
- Observed target span 2026-09-07T07:14:30Z–2026-09-10T10:59:20Z; not a prespecified evaluation window.
- Sample counts
- 56 total SOL forecast rows; 40 resolved rows.
Resolved rows by horizon: 6h n=13, 12h n=12, 24h n=9, 48h n=6. These are observed counts in the retained capture, not a complete-history claim.
Because 12h, 24h, and 48h horizons extend beyond a six-hour issue cadence, their evaluation windows can overlap. Adjacent observations should therefore not be treated as independent without row-level timestamps.
Probability, direction, and ties
A directional probability is a number from 0 to 1 (or 0% to 100%) that represents the forecast probability for the defined direction. The directional Brier formula compares that probability with a binary outcome, using p for the probability and y ∈ {0, 1} for the recorded outcome.
The supplied audit does not document the production resolution rule for an exactly unchanged price or other tie. This page therefore does not assign ties to “up” or “down,” and does not infer an undocumented tie convention.
It also does not document a production threshold for converting a probability into a directional class, so no threshold is inferred here.
Intervals, resolution, and evaluation
An 80% confidence band is a lower and upper forecast interval with 80% nominal coverage. The nominal label describes a long-run coverage target, not certainty for any individual outcome. See the dedicated 80% confidence-band coverage explanation for interval interpretation.
A resolved forecast is one whose target timestamp has passed and whose outcome price, outcome source, and scoring rule have been recorded. The retained rows include target timestamps and outcomes, but they do not establish production resolution rules, an immutable archive, or tie handling, so this page does not invent them.
The deployed card also reports a SiCierto-specific predicted-price-normalized mean absolute error percentage. Read the normalized forecast-error definition for its captured implementation and reporting boundary.
How the evaluation measures relate
The direction Brier score evaluates probabilistic direction forecasts: lower scores are better, but a score near the constant 50% reference is not proof of skill. The direction Brier-score explanation owns that baseline and interpretation.
Calibration asks whether stated probabilities or intervals agree with observed frequencies. The calibration evidence page keeps that question separate from directional correctness and price error.
Evidence boundary
The retained September 10, 2026 audit evidence contains 56 total SOL forecast rows and 40 resolved rows. Their observed target span is 2026-09-07T07:14:30Z–2026-09-10T10:59:20Z; this is row-derived context, not a prespecified evaluation window.
The figures below are a dated audit snapshot, not a live feed. The retained rows support arithmetic for the displayed Brier and coverage values, but they are point-in-time API evidence rather than an immutable research dataset. The snapshot cannot establish profitability, statistical significance, non-overlapping holdout performance, or predictive superiority.
Data provenance and export status
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.
Citation-ready public export status: NOT READY. The retained historical evidence contains 56 total SOL forecast rows and 40 resolved rows, with resolved horizon counts of 6h n=13, 12h n=12, 24h n=9, 48h n=6 and observed target span 2026-09-07T07:14:30Z–2026-09-10T10:59:20Z. These are point-in-time observations, not an immutable issued-and-resolved archive, complete history, or prespecified evaluation window. A current public feed may expose rows, but it is not an immutable versioned research archive. Metrics calculated from a changing feed are not automatically comparable to this dated snapshot. A research export would need versioned records with forecast ID, asset and venue, UTC issue and target timestamps, horizon, point prediction, interval bounds, direction probability, model version, outcome source, observed price, resolved-at time, scoring-rule version, and content hash. No research export is available from these pages.
The existing Dataset JSON-LD preserves the audit date rather than implying freshness.