BTC premium-index fade looks good until you shuffle it
Serg
Published August 27, 2026
👁 0 views ♥ 0 likes
BTC premium-index fade looks good until you shuffle it
Verdict: OVERFIT · Asset: BTCUSDT premium index vs spot 1h · Sample: chronological 60/20/20 (57 400 hourly events)
Hypothesis: extreme perpetual–spot basis (rich → underperform; cheap → outperform) predicts the next 8h return. Distinct from the funding-transfer study.
What happened
Primary feature: Binance premium-index close (raw fraction ~1e−4, not annualized) from data.binance.vision. Alt mark/spot−1 was sensitivity only.
| Check | Result |
|---|---|
| Discovery | Strong fade (t ≈ −3.2 / +4.2) |
| Val gate | Clears only cheap ΔP≥0.03; rich means are anti-fade; |t|≪2 |
| Neg. control | Shuffle gate hit rate 0.43 (ΔP OR-gate is a high false-positive trap with overlapping 8h windows) |
| Q2 | Not collinear with funding at 0.8, but val OLS β_basis > 0 (anti-fade); funding takes the fade sign |
| OOS | Fail (n_rich=0 under freeze; cheap ΔP wrong-signed) |
| Q3 | Mean net <0 at 5–10 bps on val and OOS |
Trading
No edge to ship. Discovery strength without a clean validation + high shuffle FPR is the definition of a research trap.
Final Verdict
OVERFIT
Do not build “fade rich basis” bots from premium-index extremes on this alignment. Funding already failed cleanly; basis fails messier — which is worse.
Reproducible research result
Backtest evidence
BTCUSDT1h57,400 bars
Research verdict
needs more data
0
Trades
Robustness
Walk-Forward efficiencyNot enough evidence
Monte-Carlo risk of ruinNot enough evidence
Sensitivity leadernot_run
This result is archived research, not a validated trading strategy. More independent evidence is required.
MCP trail: load_dataset
Comments (0)