H3 — Flow-Confirmed Breakout Continuation: Design (frozen)

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

Scope: H3 only. H1 (cvd_absorption_findings.md) and H2 (cvd_failed_sweeps_findings.md) are completed negatives; H4–H6 not covered.

Related: reports/breakout_regime_study.md (prior negative — see A.2).

Part A — Assessment and the overlap to beat

A.1 Reusable machinery

• Levels: analytics/sweeps.prior_range_levels (trailing prior-range high/low, excludes the current bar) — exactly the "breakout of a historical range" construct.
• Estimators: analytics/divergence (pooled_ic, decile_table, block_bootstrap_ic) and analytics/absorption (block_bootstrap, matched_spread, absorption_regression, _oriented_extremes).
• Regimes: analytics/regimes. Splits/resampling as before.
• No new data needed — perp 1m bars on disk (2026).

A.2 This is not a fresh question — the prior study already came close

reports/breakout_regime_study.md tested breakout triggers with a confluence filter cvd_20 > 0 (20-bar CVD positive) and reported validate lift 1.07, not significant (design 0.99 / validate 1.07), part of an overall negative result. H3 must therefore add something over that filter or it is a re-run:

H3's additions, all beyond the prior test:

1. Flow persistence across multiple bars (not a single cvd_20>0 sign).
2. Breakout displacement and displacement-conditioned flow.
3. Failure-rate and adverse-excursion (MAE) outcomes, not just return.
4. Explicit incremental-vs-price-only test (does flow beat a breakout alone?).
5. A wider universe (all 62 crypto perps) for power, and stated prior that a negative is likely.

The prior study is cited prominently so a null H3 is correctly read as "consistent with the earlier breakout negative", not a surprise.

A.3 Event frequency is ample (measured, 1h, 2026)

Breakout = close beyond the trailing L-bar extreme (excluding current bar):

Full power even on top-10; the 62-symbol set gives ~6× more. No tick data required. Limitation unchanged: single regime (2026) → exploratory verdict.

Part B — Frozen H3 specification

B.1 Hypothesis and rationale

Question. Do breakouts supported by persistent same-direction aggressive flow continue more than breakouts with weak flow or immediate flow reversal, after accounting for the breakout itself (price-only baseline)?

H3. Breakouts with high aligned/persistent taker flow exhibit better continuation (larger oriented forward return, lower failure rate, smaller adverse excursion) than breakouts with weak or opposing flow.

Direction. Breakout direction d = +1 (close above prior high) or d = −1 (close below prior low). Oriented forward return y_h = d · fwd_ret_h; H3 predicts y_h increases with flow alignment/persistence.

Proxy caveat. Taker flow is a proxy for genuine aggressive intent; the prior negative makes a null outcome plausible.

B.2 Definitions (causal)

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

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

bull_t = close_t > prior_high_t
bear_t = close_t < prior_low_t
d_t    = +1 if bull_t else (-1 if bear_t else 0)
break_disp_t = d_t * (close_t - level_t)/level_t      # how far beyond, signed
range_width_t = (prior_high_t - prior_low_t)/close_t  # volatility scale

delta_t   = 2*taker_buy_quote_volume - quote_volume    # signed taker flow, USDT
act_t     = mean(|delta|, trailing A)                  # activity scale
F_t       = sum(delta over [t-w+1 .. t])               # w-bar flow
flow_align_t = d_t * F_t / (w * act_t)                 # aligned, normalized
persist_t    = d_t * mean(sign(delta) over [t-w+1 .. t])  # ∈ [-1,1] signed persistence
flow_align_z_t = flow_align_t / std_W(flow_align)      # z-scored

Signals evaluated at the close of t; outcomes start at t+1. A per-symbol cooldown (primary 24 bars) de-duplicates clustered breakouts.

B.3 Required input data

Perp 1m klines (2026, on disk): OHLCV + taker-buy base/quote volume. Controls use regimes; funding optional. No spot, no tick data, no downloads.

B.4 Baselines and matched controls

1. Price-only breakout baseline — all breakouts, no flow conditioning (unconditional oriented forward return / failure rate).
2. Momentum-only baseline — condition on past return / trend only.
3. Prior-study filter — the single-bar cvd_20 > 0 condition, for direct comparison with breakout_regime_study.md.
4. Matched controls — high-persistence vs low-persistence breakouts matched on break_disp, range_width, trailing realized vol, past_ret, and regime.

B.5 Primary outcome and horizon

• Primary: oriented forward return over h = 4h (entry t+1, exit t+1+h).
• Secondary: continuation failure rate (price closes back through the breakout level within h), oriented MFE/MAE, directional hit rate.
• Robustness horizons: 1h, 24h, 3d.

B.6 Confounders and leakage

• Causal levels — exclude bar t; explicit causality test.
• Event clustering — cooldown; report whether overlaps change results.
• Volatility/range confound — control and match on range_width, realized vol.
• Trend confound — breakouts cluster in trends; control on past_ret, stratify by regime, and report the momentum-only baseline.
• Mechanical — flow is contemporaneous with the breakout 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

• Event pooled rank IC(flow_align_z, y_4h) (headline) + block-bootstrap CI; the same for persist_z.
• Tercile/quintile event study + monotonicity; failure rate by quintile.
• Matched-control spread (high − low alignment), paired within strata, CI.
• Regression: y ~ flow_align_z + break_disp + range_width + realized_vol + past_ret + regime FE, clustered by symbol×day; the flow_align_z coefficient and CI is the incremental-value test vs the price-only baseline.
• Costs: ~12 bps round-trip; a tradable claim must clear it.
• Stability: per-symbol, 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).

B.9 Feasibility with existing data

Yes. Event counts verified (A.3); all primitives either exist or are trivial to compose (prior_range_levels reused). No tick data. Binding limits: single regime (2026) → exploratory; overlap with the prior breakout negative (A.2).

Part C — Frozen primary specification

C.1 Implementation plan (after approval)

• New module src/quant_lab/analytics/breakouts.py: detect_breakouts (reuse sweeps.prior_range_levels), flow_alignment (align + persistence), failure_outcomes, run_breakout_study, robustness_sweep.
• Reuse analytics/divergence and analytics/absorption estimators; reuse regimes.
• New tests tests/test_breakouts.py: level causality, breakout direction on hand-built OHLC, flow-alignment identity, failure-outcome correctness, forward-return alignment/no-leakage, synthetic edge/no-edge fixtures.
• New script scripts/breakout_study.py; writes reports/cvd_breakout_flow_results.md + machine outputs under experiments/reports/breakouts/.
• No existing reports overwritten; no existing research code behaviour changed.

Part D — Explicit non-goals

• Not implementing H4–H6.
• Not re-running the existing breakout_regime_study triggers (bollinger_breakout etc.); H3 is a flow-conditioning event study, not a strategy backtest.
• Not tuning thresholds to find a positive result; the sweep is reported whole.
• Not using tick/aggTrades.

← All reports · source (.md)