"Edge-discovery batch 2 — v2 falsification + decomposition"

Date: 2026-09-10 | same panel/split/costs as batch 1. v2 adds: deflated Sharpe (N=5 candidates), PIT-200 universe, lag-2 execution delay, PIT-tiered liquidity costs, yearly subperiods.

#Scoreboard (base-cost OOS Sharpe | DSR | PIT-200 | lag2 | tiered | stage)

#Verdicts

#H2 liquidity-shock-reversal -> CANDIDATE

• OOS Sharpe 0.82 (IS 0.28), p vs 0 0.00, p vs BASE 1.00, corrBASE 0.32, beta 0.17
• v2: DSR 0.81 (trials-var 0.06), PIT-200 0.54, lag2 0.42, tiered 0.66, pos-years 8/11
• cost ladder: optimistic=0.93, base=0.82, conservative=0.52, severe=0.02
• decay: H5=0.56; H10=0.82; H21=0.08; H42=0.42; H63=0.24
• yearly: 2016:0.29; 2017:1.99; 2018:0.63; 2019:0.28; 2020:1.19; 2021:-0.85; 2022:-0.36; 2023:-0.34; 2024:3.34; 2025:2.95; 2026:1.08
• reason: survives v2 except severe costs (known); forward paper logging started

#H5 vol-compression-breakout -> CANDIDATE

• OOS Sharpe 0.64 (IS 0.64), p vs 0 0.02, p vs BASE 1.00, corrBASE 0.63, beta 0.59
• v2: DSR 0.69 (trials-var 0.06), PIT-200 0.40, lag2 0.74, tiered 0.54, pos-years 9/11
• cost ladder: optimistic=0.70, base=0.64, conservative=0.46, severe=0.17
• decay: H5=0.44; H10=0.61; H21=0.64; H42=0.66; H63=0.85
• yearly: 2016:0.97; 2017:1.40; 2018:0.18; 2019:0.90; 2020:-0.61; 2021:2.00; 2022:-0.53; 2023:0.59; 2024:2.39; 2025:0.70; 2026:2.33
• reason: survives v2; DSR check + holdout still required

#H6 beta-hedged-momentum -> CANDIDATE

• OOS Sharpe 1.25 (IS 0.57), p vs 0 0.00, p vs BASE 0.02, corrBASE 0.98, beta 1.05
• v2: DSR 0.91 (trials-var 0.06), PIT-200 1.06, lag2 1.23, tiered 1.22, pos-years 11/11
• cost ladder: optimistic=1.27, base=1.25, conservative=1.19, severe=1.08
• decay: H5=1.15; H10=1.17; H21=1.25; H42=1.23; H63=1.16
• yearly: 2016:1.63; 2017:2.25; 2018:0.43; 2019:1.81; 2020:1.07; 2021:1.33; 2022:0.15; 2023:1.80; 2024:2.09; 2025:1.42; 2026:1.67
• reason: holds reviewer cap: locked-holdout condition stands

#H7 overnight-leg-momentum -> RESEARCH

• OOS Sharpe 1.00 (IS 0.77), p vs 0 0.00, p vs BASE 0.60, corrBASE 0.91, beta 1.15
• v2: DSR 0.87 (trials-var 0.06), PIT-200 0.80, lag2 1.00, tiered 0.98, pos-years 10/11
• cost ladder: optimistic=1.02, base=1.00, conservative=0.96, severe=0.88
• decay: H5=0.99; H10=1.02; H21=1.00; H42=1.05; H63=1.23
• yearly: 2016:1.09; 2017:2.14; 2018:0.10; 2019:1.78; 2020:1.14; 2021:1.24; 2022:-0.29; 2023:1.09; 2024:1.63; 2025:0.98; 2026:1.48
• reason: grade A, DSR=0.87, tiered=0.98, corrBASE=0.91

#H8 intraday-leg-momentum -> RESEARCH

• OOS Sharpe 1.13 (IS 0.59), p vs 0 0.00, p vs BASE 0.66, corrBASE 0.95, beta 1.03
• v2: DSR 0.87 (trials-var 0.06), PIT-200 1.01, lag2 1.15, tiered 1.10, pos-years 10/11
• cost ladder: optimistic=1.15, base=1.13, conservative=1.06, severe=0.95
• decay: H5=1.07; H10=1.08; H21=1.13; H42=1.17; H63=1.04
• yearly: 2016:1.87; 2017:2.26; 2018:0.43; 2019:1.59; 2020:0.63; 2021:1.48; 2022:-0.14; 2023:2.12; 2024:1.38; 2025:1.39; 2026:1.85
• reason: grade A, DSR=0.87, tiered=1.10, corrBASE=0.95

Paper tracker: H2 asof 2026-08-12 gated 13 long 13; H5 asof 2026-08-12 gated 109 long 30 Paper log: logs/edge_discovery_paper.jsonl (signals only, no orders).

#Reviewer addendum — what v2 actually proved

1. H7/H8 are momentum repackaged, not new edges. corr vs BASE 0.91/0.95, p vs BASE 0.60/0.66 — close-to-close momentum is the sum of the two legs, so top-decile overlap is mechanical. The mechanical rule correctly held both at RESEARCH (DSR 0.87 < 0.9 gate). Diagnostic value only: both legs work in this universe/period (intraday OOS 1.13 slightly beats overnight 1.00, but its IS is worse, 0.59 vs 0.77 — the familiar OOS-tilt warning). Neither leg displaces BASE. No further work unless a leg-specific interaction hypothesis is articulated.
2. H2's Sharpe lives in two years. 2024 (+3.34) and 2025 (+2.95) carry the OOS case; 2021–2023 bled three straight years (-0.85/-0.36/-0.34), and lag-2 delay halves it (0.42). This is a stress-rebound edge (V-shaped vol events), not an everyday one — any sizing must survive multi-year droughts, and live fills during those rebounds will be worse than modeled. Stays CANDIDATE, paper log running, but promotion now needs forward data, not more mining.
3. DSR understates selection. N=5 covers this batch's family; cumulative trials across batches 1–2 is 8, which would shave every DSR above by a few points (H6 0.91 → ~0.88). Conclusions unchanged, but future batteries must use cumulative N.
4. H5 is the cleanest survivor. IS=OOS (0.64/0.64), delay-robust (lag2 0.74 > base), tiered-cost positive, 9/11 years. Weak only in trend-failure years (2020, 2022) — expected for a breakout leg. Still needs the locked holdout + paper record like everything else.

Runtime: 1.7 min.