"Edge-hunt36 — equity risk-appetite -> crypto daily diffusion"

SPY 2004-01-05..2026-08-27 | BTC 2013-01-01..2026-09-20 | split IS<2021-01-01/OOS>= (coins: IS<2025-01-01) | 15bps/flip | placebo = SPY lag 2..30 sessions.

Signal: prior completed SPY session return -> next crypto UTC day, long-only (flat otherwise). PIT-clean by construction.

#PRIMARY — BTC long-history legs

#Placebo (lag-shift null, PRIMARY SPY>0)

• OOS observed net SR -0.08 vs lag-null mean -0.02 / max 0.53 over 29 lags -> p 0.517
• IS observed net SR 1.16 vs null mean 0.90 -> p 0.103
• DSR(K=4) OOS net vs B&H 0.60: 0.05 | haircut -0.08

#Timing audit (why the held-window convention matters)

The first cut used the naive BTC close-to-close window [00:00 t-1, 00:00 t]. That window CONTAINS the SPY session dated t-1, so the prior-session signal sat inside the target — a lookahead. It produced a spurious OOS net SR ~3.0 and IC +0.32. The held window below ([00:00 t, 00:00 t+1]) is strictly after the session used (dated <= t-1), and the effect collapses to null, matching the clean kline legs. Any crypto day-return joined to a same-calendar equity session must use held, not close-to-close, windows.

#Dose-response (quintiles of prior SPY session ret -> next BTC ret)

#Per-year net SR (PRIMARY) vs B&H

#SECONDARY — ETH / 8-coin EW (kline cache, 2022-12+)

#Verdict: REJECT

• PRIMARY SPY>0: OOS net-vs-BH -0.68, placebo p 0.517, IS edge -0.39
• REJECT -> the SPY risk-tone overlay is indistinguishable from a stale/lagged signal; equity->crypto daily diffusion stays unproven on this sample. If PASS, graduate to shadow/log-only first (no live wiring; that needs a separate vote).