H2 — Failed Liquidity Sweeps: Findings & Recommendation

Status: complete — negative result.

Design: reports/cvd_failed_sweeps_study.md (frozen before results).

Machine outputs: experiments/reports/sweeps/{summary,by_symbol,leave_one_out,robustness}.json, events.parquet.

Reproduce: python scripts/sweep_study.py.

1. Verdict

NO EDGE (H0). All preregistered criteria Q1–Q7 fail. Aggressive flow aligned with a failed sweep does not predict the strength of the subsequent reversal.

2. What was tested

Failed sweeps of a trailing prior-range extreme (wick beyond, close back inside), 1h bars, crypto top-10, L = 48, 24-bar cooldown, 4h primary horizon; signal = activity-normalised flow aligned to the sweep direction (af_z); outcome = oriented forward return; baselines = all sweeps / low-flow sweeps / matched controls. 1,262 events (508 design, 754 test).

3. Results

• Horizons (test IC): 1h −0.0067 · 4h −0.0022 · 24h +0.0068 · 3d +0.0063.
• Matched control (high-af − low-af sweeps): −32.0 bps, CI [−125.4, +46.0] → includes zero, wrong sign.
• Control regression: af_z coefficient ≈ 0.00002, CI [−0.00070, +0.00080] → indistinguishable from zero.
• Per-symbol test IC: wildly mixed (BTCUSDT +0.152, ZECUSDT −0.172, DOGEUSDT +0.163, BNBUSDT −0.135; n≈80 each) — the spread of signs at n≈80 is what noise looks like.
• Quintiles: non-monotone.

4. Robustness sweep (18 cells)

Timeframe ∈ {15m,1h,4h} × lookback ∈ {24,48,96} × flow-window ∈ {1,4}:

• Test IC ranges −0.093 … +0.026 and design and test signs disagree across cells; no cell is stable, and the larger |IC| values sit on the smallest samples (4h, n≈200–400).
• The single "promising" design value (0.0495 at the primary spec) reverses in test — a textbook selection artefact.

5. Interpretation

The failed-sweep story (trapped aggressors → sharper reversal) is as plausible as H1's, and the event frequency is ample. The data still reject it: the design-period signal does not survive, the matched-control spread is negative and insignificant, the regression coefficient is zero, and the parameter sweep shows no stable region. This is the third independent order-flow negative on the same 2026 sample (spot–perp divergence, H1 absorption, H2 sweeps).

6. Limitations

• Single regime, in-sample 2026 — no true OOS across regimes; BTCUSDT is the only multi-year symbol.
• Event-level sample is small (1.3k events on top-10; 754 in test): low power for a modest effect. The wider 62-symbol crypto universe yields ~22k events and would materially raise power.
• Taker-flow proxy; 1h resolution; no tick data.

7. Recommendation

1. Do not re-tune H2. The sweep shows no stable cell; the design positive is noise. Rescuing it would be curve-fitting.
2. Highest-leverage step remains: build 2024–2025 perp 1m history (~2 h, ~3 GB). Three hypotheses have now been judged on a single regime; none can be confirmed without out-of-sample data.
3. If continuing bar-level flow work, widen the universe first. Event-based hypotheses (H2) are power-limited on top-10; the 62-symbol crypto set gives ~17× the events at no extra download. Re-run any survivor on the wide set.
4. Next candidate: H3 (flow-confirmed breakout continuation), on the wide universe. It is the most strategy-like, has a clean price-only baseline, and does not need new primitives. H6 (spot lead–lag) is the other distinct test, but note it risks re-treading the failed divergence study.
5. H5 as execution infrastructure. Directional price-impact asymmetry is worth keeping as a slippage/cost model even without alpha.
6. Do not buy tick data yet. The binding constraints are regime breadth and statistical power, not resolution.

Bottom line: abandon H2; the priority order is (a) out-of-sample history, (b) wider universe, (c) H3 — before any further event-level flow hypotheses.

← All reports · source (.md)