Not separated from luck

GRASSUSDT · 60m · Verdict Card

Percentile-deviation fade (mean reversion) · mr_percentile · hold 24 · 2024-11-08 → 2026-10-11 · 16849 bars · bybit · 2026-10-12 12:50

Rule: Fade when the 24-bar log deviation |d| ≥ its rolling p80 percentile (threshold recomputed every 168 bars); long & short
Cost model: 5.5bp/side fee + 3bp/side slippage + 0.07bp/bar funding

Trades

227

Win rate

58.1%

Avg / trade

110bp

Gross avg

129bp

Total

24965bp

t-stat

1.82

Permutation p

0.0190 · T0.003/S0.019

95% CI of mean

[-1bp, 231bp]

Max drawdown

69.3%

Exposure

36.7%

Profit factor

1.41

Worst trade

-3668bp

Median

198bp

Equity curve (log)

Buy & hold: -78.7% · missing 0 (max gap 0) · OHLC complete

Stability by segment

Seg	From → To	Trades	Avg / trade	Win rate	t-stat
Seg 1	2025-02-02 → 2025-07-02	57	67bp	54%	0.64
Seg 2	2025-07-11 → 2025-11-25	57	-33bp	53%	-0.24
Seg 3	2025-11-26 → 2026-05-16	57	140bp	60%	1.47
Seg 4	2026-05-21 → 2026-10-09	56	269bp	66%	1.94

In-sample/Out-of-sample	Trades	Avg / trade	Win rate	t-stat
In-sample 60%	136	41bp	56%	0.55
Out-of-sample	91	213bp	62%	2.10

Trade log · last 12 trades

Time	Side	Entry px	Exit px	Bars held	Exit	Net
2026-10-09 23:00	SHORT	0.647	0.6492	24	time	-52.6bp
2026-10-08 14:00	LONG	0.615	0.6345	24	time	293.5bp
2026-10-07 03:00	LONG	0.6826	0.6388	24	time	-681.9bp
2026-10-05 15:00	LONG	0.6955	0.7165	24	time	278.8bp
2026-10-03 22:00	SHORT	0.7565	0.6857	24	time	963.9bp
2026-10-01 20:00	LONG	0.6749	0.7058	24	time	429bp
2026-09-30 07:00	SHORT	0.7657	0.6957	24	time	940bp
2026-09-29 06:00	LONG	0.5856	0.7524	24	time	2487.6bp
2026-09-28 05:00	SHORT	0.6621	0.5718	24	time	1447.6bp
2026-09-27 04:00	SHORT	0.5796	0.6297	24	time	-847.7bp
2026-09-25 21:00	SHORT	0.4831	0.5204	24	time	-762.4bp
2026-09-24 20:00	LONG	0.4281	0.4839	24	time	1206.5bp

Method & known limits

Exit: fixed 24-bar hold; fill=signal bar close
Costs: 5.5bp/side fee + 3bp/side slippage + 0.07bp/bar funding drag (same both sides; funding sign not modeled)
Non-overlapping: signals during open positions are dropped
Significance: t-stat on per-trade net returns. Two nulls — (1) timing permutation: entry times randomized, direction mix kept, same trade count, same non-overlap suppression, p=0.0035; (2) sign-flip permutation: each trade's realized move kept, direction re-rolled (variance-matched, stricter for fades), p=0.0190. Verdict uses the stricter p=0.0190 (2000 trials, seed 42, reproducible)
The null must match the live procedure trade-for-trade: every surrogate fills the same count (227 entries, 2000 effective trials). A surrogate that only fills ~9 entries widens the null about 4x and can fake p=0.30 out of p=0.01
Uncertainty: mean 95% bootstrap CI [-1bp, 231bp] — spanning 0 means no claim of edge
Sample: 16849 bars of 60m, 0 missing (max gap 0), OHLC complete
Splits: first 60% in-sample, rest out-of-sample, plus 4 time blocks to catch single-segment flukes
Params: k=24 pct=80 side=both warmup=2000 rebalance=168 — if tuned on this same history the verdict carries selection bias; collect out-of-sample before believing it
Known limits: no order-book depth so slippage is assumed; liquidation/margin/liquidity not modeled; rejected or partial fills not modeled
Null sensitivity: the sign-flip test keeps each trade's realized move and re-rolls only the direction, so the long/short allocation itself counts as evidence — in a one-way market that can inflate significance. Hence the verdict requires both nulls to clear, a positive net mean and a positive out-of-sample mean; miss one and it degrades to "not separated from luck"

This card only judges whether a rule has statistical evidence under stated costs. Not investment advice; past results do not predict future returns.