Low-Volatility Upward Grind → Breakout Continuation — V2 (1H)

Primary question: following a bearish impulse, does a low-volatility upward grind break above its frozen high before breaking below its frozen low within 12h at a higher rate than comparable upward-drifting structures?

Executive summary

• Analysis A (boundary race): cohort A broke upward first in 461/556 resolved events (82.9%; 95% Wilson 79.6%–85.8%); all-valid denominator 83.5%.
• Comparable upward drift (no compression): 76.9%; cohort B (drift+compression, no impulse): 78.5%.
• Raw race advantage: A − drift-only = +6.1% (95% block-bootstrap CI +1.2% to +10.9%). This is the visual hypothesis on its own terms.
• Analysis B (geometry matched): pattern 82.9% vs matched 80.1% (paired diff +2.8%, 95% CI -0.4% to +6.1%); the confidence interval includes zero.
• Analysis C (equal-distance 1 ATR barriers): pattern up-first 51.3% vs matched 53.2% (paired diff -1.9%, CI -7.0% to +3.4%); both are near a coin flip.
• Equal-distance barrier level: A up-first 53.4% of 541 resolved (95% CI 49.2%–57.6%); the interval includes 50%.
• Conclusion: the raw boundary race is high and modestly above comparable upward drift, but the advantage is not robust to the boundary-distance control. Geometry-matched controls, the equal-distance barrier test, the regression controls, forward returns and the trading simulation all fail to establish a directional edge. The hypothesis is unsupported beyond geometry.

1. Audit of lowvol_grind_v1

V2 reuses the v1 detector unchanged. The previous implementation was audited against the event-level data before any change:

Reconciliation: cohort A event file (575 rows) reproduces the 12h summary exactly (461 breakout-first, 95 breakdown-first, 19 neither, 0 ambiguous); all baseline differences recompute from the saved event files. The only latent issue found was a random-control summary denominator that counted unresolved controls; it was corrected and the v1 artifact on disk is regenerated with resolved denominators.

2. Dataset and data quality

Same universe, venue and window as v1. 1m data are complete for every symbol, which makes the intrabar ordering and the strategy simulation reliable. HYPEUSDT starts 2025-05-30.

3. Pattern and structural features

The v1 primary pattern is reproduced exactly: 8-bar window; normalized log-price slope in [0.10, 0.60]; window realized-vol ratio ≤ 0.80; prior 12h impulse z ≤ −1.50; frozen pattern high/low. V2 adds structural features (slope R² and t-statistic, higher-low frequency, consecutive higher closes, candle overlap, pullback ratio, wick-to-body ratio, failed downward pushes, normalized boundary distances, time since impulse low) as candidate features.

4. Analysis A — boundary race

Breakout-first = price crosses the frozen pattern high before the frozen pattern low; 1m bars resolve same-hour touches. Rates are on resolved events (breakout-first + breakdown-first); the all-valid rate is also reported.

Pattern − comparator differences (12h, block bootstrap)

5. Analysis B — geometry-matched controls

Controls are sampled from eligible windows matched on symbol, duration, trailing volatility quintile, prior-impulse status, distance-to-high bin, distance-to-low bin, normalized-range bin and half-year, with a documented relaxation order. Matching uses detection-time information only.

Match quality

Matched-difference robustness

Geometry-conditioned linear probability models

boundary_*: breakout-first on an is-pattern indicator with geometry, volatility, slope and impulse controls (month-clustered SE). barrier_*: equal-distance up-first with the same pattern indicator (geometry is symmetric by construction).

{ "boundary_n": 2198.0, "boundary_coef_is_pattern": -0.009838015372951881, "boundary_ci_low": -0.041257028902663466, "boundary_ci_high": 0.0215809981567597, "boundary_p_value": 0.5394071259758486, "barrier_n": 2198.0, "barrier_coef_is_pattern": 0.003376220327757295, "barrier_ci_low": -0.03994343670034788, "barrier_ci_high": 0.04669587735586247, "barrier_p_value": 0.8785919221940865 }

6. Analysis C — equal-distance ATR barriers

Reference price is the first tradable price after detection (open[t+1]); barriers are ±k·ATR(trailing 14h) from that reference, for k ∈ [0.5, 1.0, 1.5, 2.0]. By construction the nearer-boundary advantage of the frozen race is removed.

A − comparator differences

7. Secondary questions

Forward returns after detection

Pattern forward returns are statistically indistinguishable from comparable upward drift at every horizon; the largest point estimate is +0.34% over 48h with a CI that includes zero.

Continuation after the breakout

Breakouts do not continue on average. In the boundary race the post-breakout 12h mean is slightly negative, 55% reverse within 3h and only 46% are still above the level after 12h. The equal-distance route is worse: 71% reverse within 3h.

8. Structural-feature ablation

Each stage adds one candidate condition to the v1 base (drift + compression + impulse). Raw breakout rates rise for some conditions, but the equal-distance up-first rate is near 50% wherever the sample remains usable, and several conditions collapse the sample to a handful of events.

9. Parameter sensitivity

Full pre-registered grid: 540 configurations. Raw-race lift over the drift-only comparator is positive in 506 of 540 cells, but the equal-distance up-first rate exceeds 50% in only 178 cells and no cell survives the multiple-testing correction on that metric. No configuration was selected using the test window.

10. Per-symbol, per-regime and chronological splits

By symbol (12h, group A)

By regime (12h)

By split (12h)

Walk-forward duration selection

11. Optional long-breakout strategy

Pre-specified long at the first trade through pattern_high, entry at the next 1m open, stop at pattern_low, 1.5R target, 12h time exit, costs of 0/3/6/12/20 bps per side plus funding. The comparison arms are a plain prior-12h-high breakout (simple) and a post-selloff-restricted version (simple_impulse).

12. Answers to the secondary questions

13. Limitations

• The geometry-matched pool is finite; exact matching on all strata is impossible and the relaxation order is documented but discretionary.
• ATR barriers at ±1 ATR are not the same object as the frozen boundaries; the two analyses answer different questions and are reported separately.
• Events overlap in time; inference uses deduplicated events and 7-day block bootstrap, but serial dependence may remain.
• The strategy is a single pre-specified rule, not an optimized system; it is secondary evidence, not proof about the pattern.

Verdict: UNSUPPORTED (raw upward rate not robust to boundary geometry)

• Boundary race: cohort A broke upward first in 461/556 resolved events (82.9%; 95% Wilson 79.6%–85.8%); all-valid rate 83.5%.
• Versus upward drift without compression: 82.9% vs 76.9% (+6.1%, 95% block-bootstrap CI +1.2% to +10.9%).
• Equal-distance ATR barriers: A up-first 53.4% of 541 resolved (49.2%–57.6%).
• Geometry-matched controls: pattern 82.9% vs matched 80.1% (paired diff +2.8%, CI -0.4% to +6.1%); equal-distance up-first 51.3% vs 53.2%.

Artifacts and reproduction

Machine-readable outputs are under experiments/reports/lowvol_grind_v2/ and figures under reports/figures/lowvol_grind_v2/. Reproduce with:

python scripts/lowvol_grind_v2_study.py

← All reports · source (.md)