"Intraday edge-discovery batch 1 — pooled event studies (1-min IEX)"

Date: 2026-09-20 | 61 names | RTH 09:30-16:00 ET | IS<2024-01-01 / OOS 2024-01-01..2025-07-01 / HOLDOUT >= 2025-07-01 | horizons (5, 10, 30, 60) min | K = 120 trials (10 hypotheses x 4 horizons x 3 variants) | cost tiers 0.5/1/2/5 bps one-way.

#Data honesty box

• Feed = Alpaca IEX 1-minute bars: OHLCV + trade_count + vwap. No NBBO, no signed flow, no depth.
• Therefore quote/order-flow families (OFI, queue imbalance, depth, cancellations, options flow) are OUT OF SCOPE — not mocked with proxies. Spread is a cost parameter, never an observation.
• VIX intraday cache is broken (empty); regime uses SPY realized vol.
• Labels are executable (enter next bar open, exit h bars later, one-bar delay) and never cross the overnight gap; alpha measured against the same-minute unconditional baseline (time-of-day control).

#Funnel

10 hypotheses tested across 4 horizons; 0 reached CANDIDATE, 10 stayed/rejected. A candidate must pass the OOS gate AND both weightings on the untouched 2025-07-01+ slice. No forward-paper claim is made here (that needs live logging).

Legend: net bps = time-of-day-adjusted forward return after round-trip cost at the base tier (1 bps one-way). day-weighted = mean of daily mean net (each session is one observation, the day-block estimator); event-weighted = mean over all events (the per-trade expectation). A real per-trade edge must be positive under BOTH — when they disagree in sign the signal is an intraday-clustering artifact (many small-count days vs a few high-count days). IS<2024-01 / OOS 2024-01..2025-07 (validation) / HOLDOUT >=2025-07 (untouched). DSR discounted by K trials.

#Alpha decay (event-weighted OOS net bps by horizon)

#Per-hypothesis detail

#C1 — ilq-extreme-move-reversal (stage RESEARCH)

• Mechanism: Transient liquidity imbalance / aggressive flow overshoots; market-makers are compensated for absorbing it.
• Hypothesis: A 5-min move beyond +/-2 intraday-sigma reverts over the next 5-60 min.
• Params: {'family': 'intraday-liquidity-reversal', 'new_mechanism': True, 'thr': 2.0}
• Decision reason: p=1.000 fails 0.05
• Perturbation OOS net bps (best across horizons per variant): ['-1.913', '-1.308']
• OOS net bps: day-weighted -1.832 vs event-weighted -1.854 (IS: -1.708 vs -1.917)
• HOLDOUT net bps: day-weighted -1.793 vs event-weighted -1.946 (301 days)
• OOS decomposition: n/a (event-weighted total not positive)
• Positive OOS years: 0/2 (DSR 0.00, haircut -4.40)

#C2 — ilq-extreme-move-vol-shock (stage RESEARCH)

• Mechanism: Volume confirms the move was inventory/flow-driven, not information; the pressure resolves once the flow passes.
• Hypothesis: Extreme 5-min moves ON abnormal volume revert harder than quiet ones (forced-flow exhaustion).
• Params: {'family': 'intraday-liquidity-reversal', 'new_mechanism': True, 'thr': 2.0, 'vzmin': 1.5}
• Decision reason: p=0.923 fails 0.05
• Perturbation OOS net bps (best across horizons per variant): ['-1.696', '+0.228']
• OOS net bps: day-weighted -1.150 vs event-weighted -1.153 (IS: -0.985 vs -2.298)
• HOLDOUT net bps: day-weighted -2.112 vs event-weighted -3.270 (301 days)
• OOS decomposition: n/a (event-weighted total not positive)
• Positive OOS years: 0/2 (DSR 0.00, haircut -1.14)

#C3 — ilq-extreme-move-range-widen (stage RESEARCH)

• Mechanism: Range blow-out marks liquidity withdrawal; the dislocation closes when depth returns.
• Hypothesis: Extreme moves while the bar range widens (spread-widening proxy) revert.
• Params: {'family': 'intraday-liquidity-reversal', 'new_mechanism': True, 'thr': 2.0, 'rzmin': 1.5}
• Decision reason: p=0.922 fails 0.05
• Perturbation OOS net bps (best across horizons per variant): ['-1.487', '+0.050']
• OOS net bps: day-weighted -1.047 vs event-weighted -0.890 (IS: -0.261 vs -1.851)
• HOLDOUT net bps: day-weighted -2.292 vs event-weighted -3.455 (301 days)
• OOS decomposition: n/a (event-weighted total not positive)
• Positive OOS years: 0/2 (DSR 0.00, haircut -1.14)

#I1 — vshock-continuation (stage RESEARCH)

• Mechanism: Abnormal volume marks informed/institutional initiation; the position is built over several minutes.
• Hypothesis: Large 5-min moves on >2z volume CONTINUE over 5-60 min.
• Params: {'family': 'intraday-volume-shock', 'new_mechanism': True, 'thr': 2.0, 'vzmin': 2.0}
• Decision reason: p=0.971 fails 0.05
• Perturbation OOS net bps (best across horizons per variant): ['-1.849', '-1.830']
• OOS net bps: day-weighted -2.250 vs event-weighted -0.041 (IS: -3.469 vs -2.131)
• HOLDOUT net bps: day-weighted -2.573 vs event-weighted -0.760 (301 days)
• OOS decomposition: n/a (event-weighted total not positive)
• Positive OOS years: 0/2 (DSR 0.00, haircut -1.64)

#I2 — vshock-reversal (stage RESEARCH)

• Mechanism: Flow-driven crowding; the liquidity premium is paid to whoever fades the exhausted impulse.
• Hypothesis: Large 5-min moves on >2z volume REVERSE (competing leg).
• Params: {'family': 'intraday-volume-shock', 'new_mechanism': True, 'thr': 2.0, 'vzmin': 2.0}
• Decision reason: p=0.884 fails 0.05
• Perturbation OOS net bps (best across horizons per variant): ['-1.527', '+0.941']
• OOS net bps: day-weighted -1.017 vs event-weighted -0.882 (IS: -0.794 vs -2.329)
• HOLDOUT net bps: day-weighted -2.280 vs event-weighted -3.711 (301 days)
• OOS decomposition: n/a (event-weighted total not positive)
• Positive OOS years: 1/2 (DSR 0.00, haircut -0.94)

#J1 — volregime-shock-continuation (stage RESEARCH)

• Mechanism: Vol expansion accompanies information arrival; trend strategies and stops chase the new state.
• Hypothesis: After an intraday volatility-regime shock, the prevailing 5-min direction continues.
• Params: {'family': 'intraday-vol-regime', 'new_mechanism': True, 'rvzmin': 2.0, 'zmin': 0.5}
• Decision reason: p=0.981 fails 0.05
• Perturbation OOS net bps (best across horizons per variant): ['-1.853', '-1.073']
• OOS net bps: day-weighted -1.616 vs event-weighted -2.159 (IS: -3.638 vs -2.721)
• HOLDOUT net bps: day-weighted -3.929 vs event-weighted -3.156 (298 days)
• OOS decomposition: n/a (event-weighted total not positive)
• Positive OOS years: 0/2 (DSR 0.00, haircut -1.76)

#J2 — volregime-shock-reversal (stage RESEARCH)

• Mechanism: Vol spikes are liquidity events; the impulse overshoots and mean-reverts once the vol shock decays.
• Hypothesis: After an intraday volatility-regime shock, the move reverses (competing leg / overreaction).
• Params: {'family': 'intraday-vol-regime', 'new_mechanism': True, 'rvzmin': 2.0, 'zmin': 0.5}
• Decision reason: p=1.000 fails 0.05
• Perturbation OOS net bps (best across horizons per variant): ['-1.777', '-1.617']
• OOS net bps: day-weighted -1.760 vs event-weighted -1.956 (IS: -1.462 vs -1.818)
• HOLDOUT net bps: day-weighted -1.890 vs event-weighted -2.004 (299 days)
• OOS decomposition: n/a (event-weighted total not positive)
• Positive OOS years: 0/2 (DSR 0.00, haircut -4.96)

#H1 — vwap-stretch-fade (stage RESEARCH)

• Mechanism: VWAP is the institutional execution benchmark; execution algos lean against extension and draw price back to it.
• Hypothesis: Price stretched >2 vol-units from session VWAP reverts toward VWAP over 5-60 min.
• Params: {'family': 'intraday-vwap-state', 'new_mechanism': True, 'thr': 2.0}
• Decision reason: date-weighted +7.36 but event-weighted -0.66bps <= 0 (clustering artifact)
• Perturbation OOS net bps (best across horizons per variant): ['+4.297', '+6.852']
• OOS net bps: day-weighted +7.364 vs event-weighted -0.659 (IS: +5.367 vs -4.260)
• HOLDOUT net bps: day-weighted +3.267 vs event-weighted -2.752 (301 days)
• OOS decomposition: n/a (event-weighted total not positive)
• Positive OOS years: 2/2 (DSR 0.00, haircut 1.86)

#M1 — relative-dislocation-reversal (stage RESEARCH)

• Mechanism: Idiosyncratic intraday dislocations without news are liquidity noise; index/stat-arb flow pulls them back.
• Hypothesis: A stock's 30-min move that dislocates from SPY's same-window move reverts to the market's path.
• Params: {'family': 'intraday-cross-asset', 'new_mechanism': True, 'thr': 2.0}
• Decision reason: date-weighted +2.23 but event-weighted -1.34bps <= 0 (clustering artifact)
• Perturbation OOS net bps (best across horizons per variant): ['-0.119', '+4.346']
• OOS net bps: day-weighted +2.235 vs event-weighted -1.335 (IS: +1.619 vs -2.375)
• HOLDOUT net bps: day-weighted +0.401 vs event-weighted -2.569 (127 days)
• OOS decomposition: n/a (event-weighted total not positive)
• Positive OOS years: 2/2 (DSR 0.00, haircut 0.00)

#E1 — spy-lead-laggard (stage RESEARCH)

• Mechanism: Index/basket flow hits SPY first; single-name liquidity refreshes slower, so the constituent completes the move.
• Hypothesis: When SPY jumps >2z in 5 min, a large cap that has not followed (gap>10bps) catches up over 5-60 min.
• Params: {'family': 'intraday-cross-asset', 'new_mechanism': True, 'thr': 2.0, 'gapmin': 0.001}
• Decision reason: p=0.983 fails 0.05
• Perturbation OOS net bps (best across horizons per variant): ['-1.349', '-1.539']
• OOS net bps: day-weighted -1.715 vs event-weighted -0.461 (IS: -1.444 vs +0.027)
• HOLDOUT net bps: day-weighted -0.479 vs event-weighted +1.089 (127 days)
• OOS decomposition: n/a (event-weighted total not positive)
• Positive OOS years: 0/2 (DSR 0.00, haircut -1.72)

Runtime: 14.8 min. Ledger: logs/edge_discovery_ledger.jsonl.