Evidence · calibration

A coverage record you can inspect.

A frozen held-out evaluation and a rolling production audit answer different questions. The first asks whether a method worked on a fixed test set. The second asks whether served bands are still behaving as claimed as new outcomes settle. We publish both—and we display a sequential breach rather than blending it into an old headline number.

Loading…
Rolling service record · 5-day band
This is a dated, rolling record of past predicted-versus-realized coverage, not a guarantee and not evidence of trading edge. Inspect the full receipt.

Two records, two questions

RecordWhat it answersHow to use it
Rolling service recordAre settled bands currently achieving nominal coverage, overall and by condition?Check the map date, sequential verdict, earnings proximity, and the slice matching the subject.
Frozen held-out evaluationDid a specific method restore coverage on a fixed, disjoint test protocol?Treat it as historical method evidence, never as the current service status.
Frozen method finding · April 2026

Why the bands needed calibration

Nearest-neighbor quantiles under-covered on the frozen evaluation because neighbors are selected for similarity rather than sampled randomly. Split-conformal offsets widened those historical bands toward nominal coverage. These are dated validation results, not today’s rolling receipt.

Frozen bandNominalRaw empiricalCalibrated empirical
[p10, p90], 5d80%~68%~82%
[p10, p90], 10d80%~68%~80%
[p25, p75], 5d50%~40%~49%

Consumption rules

  • — Check the live receipt before relying on nominal coverage. A sequential breach means investigate or abstain; it is not a cosmetic warning.
  • — Do not use the standard band unchanged near earnings. Use the published proximity correction or avoid the uncertainty claim.
  • — A marginal band for one horizon is not a joint path guarantee. Use the hold-path band when the entire 1d-to-10d path must remain inside.
  • — Calibration measures uncertainty honesty. It does not establish direction, expected return, implementability, or trading edge.