Learned regime-conditional signal-combination meta-model

The question

The research methodology itself — the pre-registration rigor protocol applied to every hypothesis below.

How it was tested

The hypothesis, its exact trigger, the outcome it predicts, and the pass/fail bar were written down and cryptographically hash-locked before any data was tested, then held through a mandatory cooling-off period. The backfill ran exactly once against those frozen parameters — no re-tuning, no curve-fitting — using adversarial statistics (stationary-bootstrap confidence intervals, multiple-testing correction, purged cross-validation with an embargo, and walk-forward out-of-sample splits). The specific trigger thresholds are proprietary and omitted here.

The outcome

RAN 2026-06-30 (attended) → DATA-LIMITED / INCONCLUSIVE — NOT a NOISE result, the test could not be FORMED. The locked methodology (purged 5-fold + 21d purge + 21d embargo) needs a per-(date,4-strategy) SIGNAL PANEL over ≥2-3yr, but the only such panel (`tepper_sleeve_signals`: strategy/signal_value/ic_estimate/net_alpha_bps/regime) spans just 2026-03-29→06-30 (~63 trading days) — far too short to form even one valid purged fold (21d purge+embargo alone exceeds the per-fold test window). The 5yr reconstructions the pre-reg assumed would extend the panel are AGGREGATE-STATS artifacts (pead_5yr = {stats,sample}; the only aligned one is pead_eqmom_monthly_aligned = 2-strategy MONTHLY RETURNS, not signals), NOT per-date panels. Running a shorter-window split would violate Rule #5 (unauthorized methodology change) → not done; no overfitting-prone fit was attempted. Live portfolio.ts IC-weighting UNCHANGED, zero trade-path. Forward path (Rule #10-clean amendment): a true per-date 4-strategy signal-panel reconstruction OR the pre-reg's binding-forward path (wait for live multi-strategy history to reach CV-sufficient length) — a data-access decision, not a retune. `data/backfill/learned_signal_combination_2026-06-30.json`. Run-once honored. — LOCKED 2026-06-22 (.1% evaluation — the 'no learned combination' gap). First LEARNED signal-combination / first ML for portfolio construction (vs hand-specified IC-weighting). The whole test is the no-subsumption-vs-IC-weighted bar: the learned model must BEAT the static baseline OOS net-of-cost or it's NOISE. Overfitting is the central risk → stricter discipline: mandatory regularization, feature-count cap (no mining), in-fold nested CV only, shuffled-label negative control, purged-5fold+embargo. Honest prior INFO_GRADE-leaning — a NOISE result (the IC-weighting wins) is a VALUABLE vindication of the simple combiner against ML hype. Live portfolio.ts combiner NEVER modified until a verdict + explicit operator promotion. Harness scripts/backfill-learned-signal-combination.ts gated. DO-NOT-RUN before 2026-06-27.

The backfill ran and concluded — the hypothesis did not fail, but the result is informational rather than a clean strategy-grade pass. See the outcome below.