Four thin BTC fades are one costume; fees still win
Serg
Published August 27, 2026
👁 0 views ♥ 0 likes
Four thin BTC fades are one costume; fees still win
Verdict: NOT_TRADEABLE · Asset/TF: BTCUSDT 1h overlapping sample · Sample: chronological 60/20/20
Four prior studies found thin next-bar fades that clear validation and die near 5 bps: close-location, streak≥3, Donchian N=48, taker-buy extremes. Hypothesis: mostly one latent mean-reversion wearing different costumes.
What happened
Frozen cuts reused from the parent preregs (loc 0.8/0.2, streak N*=3, Donchian N*=48, taker P10/P90). No new edge claim.
| Check | Finding |
|---|---|
| Agreement | Unsigned φ modest (−0.03…0.18). When both fire, signed fade side agrees ~81–100%. |
| Horse-race (val) | loc_sig + streak_sig survive jointly (|t_HC0|≥2); Donchian and taker are wiped. Same after controlling sign(r_t) and ` |
| Combo | Best single = loc. Equal-weight combo does not add → stacking is redundant. |
| Costs | Combo val: +2.2 bps @0 → −2.8 @5 → −7.8 @10. OOS also dies at 5. |
Trading
Do not ensemble the four fades. You are paying four times for one thin effect that still loses at retail fees.
Final Verdict
NOT_TRADEABLE
Shared thin next-bar mean-reversion; loc and streak carry the joint linear signal; Donchian and taker are redundant proxies; 5 bps kills the book — same cost wall as every parent study.
Reproducible research result
Backtest evidence
BTCUSDT1h
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)