Methodology changelog

Current public methodology version: v1.2.1

When grading rules change, we record it here so citations can pin a version. Full rules live on the accuracy methodology page.

  1. v1.2.1 · 2026-08-26

    Honesty & transparency: insufficientSample gates, secondary post-outage / post-1.2.0 cohorts, never remap HOLD-overall into directional; ops dark-window alert.

    • Short maturity windows with fewer than 30 directional grades are marked insufficientSample and are not citeable.
    • /accuracy.json exposes cohorts.post_outage and cohorts.post_methodology_120 beside the ALL headline.
    • Public snapshot never copies overall (incl. HOLD) into directionalAccuracy when directional n=0.
    • Cache-miss path refuses short maturity timeframes (requires rebuilt cache).
    • Ops SLO alerts when PredictionHistory has zero rows in the last ~2.5 days (prediction_dark_window).
    • Force 1D regrade: POST /api/accuracy/full-pipeline body.forceReeval + scripts/accuracy-force-1d-then-cache.ps1.

  2. v1.2.0 · 2026-08-25

    1D (and graded horizons) use next US equity session closes; fuzzy lookup prefers forward prices and never collapses to the base session (fixes weekend ~0% returns that crushed directional accuracy).

    • 1D target = next NYSE/Nasdaq trading session after the signal’s Eastern calendar day (Fri → Mon, not Sat).
    • Longer horizons: calendar offset snapped forward onto a trading day.
    • Close resolution order: exact → forward sessions → prior sessions; never the base session close.
    • Rows that only resolve to the base day are left unevaluated (no fake 0% grades).
    • Re-grade 1D after deploy with HOUND_ACCURACY_FORCE_REEVAL=1d then rebuild accuracy-cache.json.

  3. v1.1.0 · 2026-03-01

    Public headline metric switched to directional accuracy (BUY+SELL only) at the 1-day horizon.

    • Directional hit rate is the primary public metric; overall (including HOLD) is secondary.
    • HOLD remains graded with horizon bands; do not mix HOLD into the directional headline.
    • Advanced/consensus subset reported separately when sample size allows.
    • Machine-readable snapshot fields: directionalAccuracy, directionalTotal, methodologyVersion.

  4. v1.0.0 · 2025-11-01

    Initial public forward-test methodology against DailyClose outcomes.

    • Signals graded after live market outcomes (not backtest-only marketing).
    • BUY/SELL correct if price moves in signal direction; HOLD correct within band.
    • Multiple horizons tracked (1D through 1Y); 1D used for public communication.

Subscribe via /accuracy-feed.xml. Also see limitations, accuracy.json, and the disclaimer.