Skip to content

Rule significance testing — is a family's edge distinguishable from zero at the fee actually paid? #475

Description

@eaitbrahim

Part of the research suite — sequenced to land before the 2026-09-30 pooled review (#427, discussion #359).

What

Per-rule-family significance testing on the cached candles: is this family's measured edge distinguishable from zero at the fee actually paid? It sits beside PBO/CSCV (keel/research/cscv.py), the DCA benchmark, and Monte Carlo resampling (#441) — #441 asks "is the equity curve an outlier?", this asks "is the edge real at all?"

Framing (non-negotiable)

Report-only evidence for the promotion gate. The honest result stands: no shipped family is net positive at the 120 bp taker fee, and inside the fee-free allowance the reconstruction sits at break-even (14.9% win vs 14.88% required). This tool must be able to say "no" — that is its job.

Acceptance

  • A significance artifact per family (CLI or provenance-tracked notebook), with the fee regime stated
  • Power/CI stated with the n_eff correction (DEFF 2.58; herding ICC 0.21) — never raw n
  • Tests first: synthetic edges the tool must detect, and nulls it must refuse to bless

Activity

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

Labels

No labels
No labels

Type

No type

Projects

No projects

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions