One-shot test of the last untested external-data hypothesis: suppress trend
entries when perp funding is elevated (crowded longs). Data: Binance ETHUSDT
8h funding, 2021-01-01 → 2026-08-17 (data/raw/external/binance_vision/ funding_ETHUSDT.csv), full coverage of the phase-13 window. Causal features
(trailing mean of last K stamps at close[t]), fixed absolute thresholds,
exits untouched, maker fees, next-open fills. Script:
scripts/funding_overlay_test.py.
Baseline reproduced exactly: +15,799 bps / +258.9% / DD −17.4% (cal +4,860 / val +10,938) — harness sanity confirmed against phase 13.
| variant | net | compounded | maxDD | cal | val | suppressed-trade PnL |
|---|---|---|---|---|---|---|
| baseline | +15,799 | +258.9% | −17.4% | +4,860 | +10,938 | — |
| K=24h, thr 1bp/8h | +11,658 | +143.9% | −24.3% | +5,646 | +6,012 | +1,261 (11 tr) |
| K=24h, thr 3–5bp | = baseline | 0 entries blocked | ||||
| K=72h, thr 1bp/8h | +11,338 | +135.9% | −24.3% | +5,405 | +5,933 | +1,483 (5 tr) |
| K=72h, thr 3–5bp | = baseline | 0 entries blocked |
Verdict
REJECT — the hypothesis is inverted in this data. At the only threshold that actually blocks entries (1 bp/8h ≈ 11% APR), the suppressed trades were net profitable (+1,261 / +1,483 bps): elevated funding at trend-onset marks strong trends, not exhausted ones. Gating those entries cuts net by ~4,100 bps and deepens maxDD (−17.4% → −24.3%). Higher thresholds are no-ops (block zero entries).
This closes the external-data question for crypto: funding/OI/LS/supply data adds nothing as signal (phase 13 §5.7), nothing as entry-risk overlay (this test). Remaining non-predictive uses are monitoring only.