← all research
REJECTED

"Shock-Day Entry Filter on gate60 (price-only, daily)"

2026-08-26

Window: 2023-01-25 .. 2026-07-31 (1284 days, 352 gate60-long days) | features: |close->close ret| (1d, 2d), range/open, overnight gap | shifted +1d | gates on long days only, gross

VariantDaysLongDaysRetSharpePFMaxDD
baseline (gate60)1284352+181.60%1.031.33-26.71%
skip abs1>=0.0201284185+66.81%0.751.33-26.84%
skip abs1>=0.0251284216+87.98%0.851.35-28.79%
skip abs1>=0.0301284240+226.29%1.351.59-19.41%
skip abs1>=0.0401284280+115.79%0.861.31-31.23%
skip abs1>=0.0501284302+159.67%1.011.36-32.37%
skip abs1<=0.0101284242+102.85%0.831.32-25.94%
skip abs2>=0.051284259+75.84%0.691.25-24.86%
skip abs2>=0.081284311+120.30%0.871.30-28.58%
skip rng>=0.061284249+100.14%0.831.32-19.82%
skip rng>=0.091284321+146.66%0.951.32-22.62%
skipgap>=0.021284352+181.60%1.03
skipgap>=0.031284352+181.60%1.03
skip down-shock only (<-0.03)1284305+217.71%1.231.44-19.99%
skip up-shock only (>+0.03)1284287+189.21%1.111.41-25.69%

gate60-long-day forward returns by prior-shock bucket

| prior |ret| bucket | Mean daily ret | Days | |---|---|---| | <1% | +0.348% | 110 | | 1-2% | +0.295% | 75 | | 2-3% | +1.293% | 55 | | 3-5% | -0.299% | 62 | | >5% | +0.221% | 50 |

Overlap audit vs vol-level family (rejected 2026-08-25)

IS/OOS split at 2025-01-01 (rule picked by IS Sharpe, frozen before OOS)

Top-5 rules on IS:

SegmentVariantLongDaysRetSharpePFMaxDD
ISbaseline256+86.70%1.061.29-26.71%
ISskip abs1>=0.030184+131.71%1.581.57-15.45%
OOSbaseline96+50.83%0.991.40-18.42%
OOSskip abs1>=0.03056+40.82%1.021.61-15.21%

OOS gated long-days: 56 | mean +0.6907% | bootstrap p(gate adds nothing): 0.656 | NULL-R permutation p (2000): 0.073

Rule neighbors (abs1 family), full window

VariantLongDaysRetSharpePF
skip abs1>=0.020185+66.81%0.751.33
skip abs1>=0.025216+87.98%0.851.35
skip abs1>=0.030240+226.29%1.351.59
skip abs1>=0.040280+115.79%0.861.31
skip abs1>=0.050302+159.67%1.011.36
skip abs1<=0.010242+102.85%0.831.32

Confound controls (full window)

Control variantLongDaysRetSharpePF
control: skip if prior-day ret<0 (reversal-only)201+6.69%0.191.07
winner OR prior-down (is winner just reversal?)287+189.21%1.111.41
winner net of top-quintile RV10 (event vs level)307+166.79%1.031.36
half size when flagged352+208.67%1.231.42

Per-year: baseline vs winner

YearBase retGated retBase LDGated LD
2023+26.35%+23.79%6654
2024+48.86%+69.05%190130
2025+48.80%+38.68%9656
2026+0.00%+0.00%00

Cross-coin: same frozen rule via vectorized twin (own SMA200/ROC60 + BTC ROC60)

CoinVariantLongDaysRetSharpePFMaxDD
BTCbaseline545+157.67%1.131.28-27.21%
BTCskip abs1>=0.030396+145.94%1.261.39-18.89%
SOLbaseline402+268.65%1.021.27-56.29%
SOLskip abs1>=0.030290+169.56%0.931.29-40.10%

Decision memo

VERDICT (2026-08-25): REJECT — family to ledger (with one caveat)

Criterion-by-criterion:

  1. OOS: gated +40.8%/Sharpe 1.02 vs baseline +50.8%/0.99 — total return DOWN, risk-adjusted flat. The gate skipped 40 long days that summed to +10.0% (mean +0.25%): on OOS it removed net WINNERS; per-day density improved (+0.69% vs +0.53%) but a fully-invested sleeve earns nothing from cash days.
  2. Significance FAILS: bootstrap p=0.656; permutation p=0.073 misses the <0.05 bar.
  3. Knife-edge neighbors: full-window Sharpe by threshold 0.85 / 1.35 / 0.86 / 1.01 at 2.5/3/4/5% — the 3% cell is an isolated spike.
  4. Bucket table non-monotone: the "effect" is only 2-3% (+1.29%) vs 3-5% (-0.30%) while >5% shock days are fine again (+0.22%). No mechanism.
  5. Vol-level netting erases it: "winner net of top-quintile RV10" falls to +166.8%/Sharpe 1.03 ≈ baseline — i.e. it mostly re-labels the high-vol regime just rejected in vol_regime_gate60.md.
  6. Cross-coin inconsistent: BTC mildly better (Sharpe 1.13->1.26), SOL clearly worse (ret +269%->+170%, Sharpe 1.02->0.93).

Caveat worth one line in history: permutation p=0.073 is the strongest round-2 signal so far and OOS PF improved (1.61 vs 1.40). If anyone revisits, it needs a NEW mechanism hypothesis (e.g. shock direction conditioned on gate-cross freshness, or half-size-only deployment) — NOT another threshold grid on abs1. Threshold re-grids are ledger-banned.