H2 — Failed Liquidity Sweeps: Design (frozen)

Status: frozen design — awaiting go-ahead before implementation.

Scope: H2 only (failed sweeps). H1 is a completed negative (reports/cvd_absorption_findings.md); H3–H6 are not implemented here.

Related: reports/cvd_absorption_study.md, reports/spot_cvd_divergence_study.md.

Part A — What H2 adds, and what must be built

A.1 The one missing capability

The repo has no swing / pivot detection (verified: grep for swing|pivot|sweep|breakout finds only the four breakout strategies, none of which detect structural levels causally). H2 therefore must build a causal level detector as its first component. Everything else is reusable: resample_bars, read_dataset, analytics/regimes, analytics/divergence estimators (pooled_ic, decile_table, block_bootstrap_ic), and the absorption.py helpers (matched_spread, block_bootstrap).

A.2 Level definition: trailing extreme (primary), pivot (robustness)

Two candidate "previously established level" definitions:

1. Prior range extreme (primary). As of bar t, the level is the highest high (for a swing high) over the trailing L bars strictly before t: prior_high_t = max(high[t-L .. t-1]). This needs no pivot detection, has a single parameter L, is unambiguously causal, and is the standard "N-bar-high failed breakout" construct. Symmetric for prior_low_t.
2. Confirmed fractal pivot (robustness). A pivot high at p is confirmed once k bars have closed without exceeding it; at time t only pivots with p + k ≤ t are usable. More literal to "swing high", but adds a knob k and a sparser event set.

Primary = (1) because it is look-ahead-free by construction and knob-light; (2) is reported as a robustness variant.

A.3 Event frequency is sufficient (measured)

1h bars, 2026-01…09, prior-range definition, count of wick-beyond-then-close-inside:

So even the top-10 yields ~3.6k events at L=48 — ample, and the wider crypto set gives ~22k. No tick data required. The only limitation remains regime breadth (2026-only; BTC is the sole multi-year symbol).

Part B — Frozen H2 specification

B.1 Hypothesis and economic rationale

Question. When price sweeps a previously established swing high/low and closes back inside the prior range (a failed sweep / stop-run), does the magnitude and direction of aggressive flow during the sweep predict the strength of the subsequent reversal?

H2. A failed sweep with large aggressive flow in the sweep direction (chasers buying an upside sweep, or selling a downside sweep) reverses harder than a similar failed sweep without such flow. Rationale: trapped aggressors and absorbed liquidity produce an outsized move once the sweep fails. Proxy, not mechanism — we cannot see stopping/absorption identities.

Directional definition (fixed).

• Upside sweep (price wicks above the prior high, closes back below) → predicted reversal d = −1 (down).
• Downside sweep (wicks below prior low, closes back above) → d = +1 (up).
• Aligned flow af_t = −d_t · F_t (flow in the sweep direction is positive): upside → net buying positive; downside → net selling positive.
• Oriented forward return y_t = d_t · fwd_ret_t. H2 predicts y_t increases with af_t.

B.2 Event and signal definitions (causal)

Per symbol, on 1h bars (resampled from 1m):

prior_high_t = max(high[t-L .. t-1])        # trailing, excludes t
prior_low_t  = min(low[t-L .. t-1])

upside_sweep_t  = high_t > prior_high_t and close_t < prior_high_t
downside_sweep_t= low_t  < prior_low_t  and close_t > prior_low_t

delta_t   = 2*taker_buy_quote_volume - quote_volume          # signed taker flow, USDT
F_t       = sum(delta over the sweep bar)   (primary: the sweep bar only)
act_t     = mean(|delta|, trailing A)                        # activity scale
af_t      = (-d_t) * F_t / act_t                             # aligned, normalized
af_z_t    = af_t / std_W(af)                                 # z-scored
sweep_depth_t = (high_t/prior_high_t - 1)  or (1 - low_t/prior_low_t)
range_width_t = (prior_high_t - prior_low_t) / close_t

Signals are evaluated at the close of t; outcomes start at t+1. The level excludes bar t (no look-ahead in level formation).

B.3 Required input data

Perp 1m klines (2026, already on disk): timestamp, open, high, low, close, volume, quote_volume, taker_buy_volume, taker_buy_quote_volume. Controls use regimes (computed from bars) and, where available, funding (100 symbols). OI is not required (partial coverage would only limit an optional robustness slice).

B.4 Baselines and matched controls

1. All failed sweeps (no flow conditioning) — unconditional oriented return.
2. Low-flow failed sweeps — the natural control group (bottom tercile af).
3. Successful breakouts — close beyond the level (not a failed sweep); tests whether the failure is what matters, not the wick per se.
4. Matched controls — high-af vs low-af failed sweeps matched on sweep_depth, range_width (ATR-normalised), trailing realized vol, trailing return and regime.

B.5 Primary outcome and horizon

• Primary: oriented forward return over h = 4h (entry close t+1, exit close t+1+h) on failed-sweep events, pooled across symbols.
• Secondary: oriented MFE/MAE, directional hit rate, time-to-reversal.
• Robustness horizons: 1h, 24h, 3d (reported in full).

B.6 Confounders and leakage

• Look-ahead in levels — prior_high/low exclude t; pivots used only after confirmation. Explicit causality test.
• Event clustering — de-duplicate with a cooldown of C bars per symbol (primary C = 24; one event per day), so overlapping sweeps are not double-counted; also dedupe consecutive bars.
• Volatility / range confound — control and match on range_width, realized vol, ATR.
• Trend confound — a failed low-sweep in an uptrend is not symmetric to one in a downtrend; control for trailing return and stratify by regime.
• Mechanical — the flow is contemporaneous with the sweep bar (known at close); forward window starts t+1.
• Multiple testing — one primary spec; sweep reported whole with BH-FDR.

B.7 Statistical evaluation & cost assumptions

• Pooled rank IC(af_z, y_4h) on events (headline), block bootstrap CI (blocks ≈ horizon; also cluster-aware via timestamp blocks).
• Tercile/quintile event study of y_4h by af_z, monotonicity check.
• Matched-control spread (high-af − low-af), paired within strata, CI.
• Regression: `y ~ af_z + sweep_depth + range_width + realized_vol + past_ret
  • regime FE, clustered by symbol×day; report the af_z` coefficient + CI.
• Hit rate of predicted reversal direction; event counts and frequency.
• Costs: ~12 bps round-trip; a tradable claim must clear it.
• Stability: per-symbol and per-regime; chronological design/test.

B.8 Pass/fail criteria (preregistered)

Declare an edge iff ALL hold on the test period:

Anything short of all seven → H0 (no edge), reported as a negative result.

B.9 Feasibility with existing data

Yes. 1m perp bars suffice (event counts verified, §A.3). No tick data, no new downloads. The binding limitations are: (i) the level detector must be built and causality-tested, and (ii) single regime (2026) → exploratory verdict, with BTCUSDT as the only multi-year cross-check.

Part C — Frozen primary specification

C.1 Implementation plan (after approval)

• New module src/quant_lab/analytics/sweeps.py: prior_range_levels, detect_sweeps, aligned_flow, sweep_signal, matched_spread (reuse from absorption), run_sweep_study.
• Reuse analytics/divergence.py estimators and analytics/absorption.py (block_bootstrap, matched_spread); reuse regimes.
• New tests tests/test_sweeps.py: level causality (level excludes bar t), sweep detection correctness on hand-built OHLC, cooldown de-duplication, forward-return alignment/no-leakage, synthetic edge and no-edge fixtures.
• New script scripts/sweep_study.py; writes reports/cvd_failed_sweeps_results.md
  • machine outputs under experiments/reports/sweeps/.
• No existing reports overwritten; no existing research code changed.

Part D — Explicit non-goals

• Not implementing H3–H6 in this pass.
• Not tuning thresholds to find a positive result; the sweep is reported whole.
• Not using tick/aggTrades.
• Not claiming cross-regime generality from a 2026-only sample.

← All reports · source (.md)