Low-Volatility Upward Drift Breakdown Study (1H)

Research question (frozen): following a bearish impulse, does a low-volatility, gently upward-trending consolidation break below its frozen low before its frozen high within 12 hours at a materially higher rate than comparable consolidations?

Executive summary

• Primary cohort A: 95 of 556 resolved events (575 deduplicated events displayed) broke down first within 12h: 17.1% (95% Wilson CI 14.2%–20.4%; all-valid denominator 16.5%); breakout-first 82.9%; neither 3.3%; ambiguous 0.00%.
• Generic low-volatility consolidation baseline (no positive-drift requirement): 51.0% breakdown-first. Pattern minus baseline = -33.9% (95% block-bootstrap CI -38.3% to -29.3%).
• Horizon curve (cohort A): 3h 10.3% · 6h 15.0% · 12h 17.1% · 24h 18.5% · 48h 18.8%.
• 12h baselines: low_vol_all 51.0%; low_vol_sideways 49.3%; drift_no_low_vol 23.1%; impulse_drift 26.3%; post_impulse_all 53.4%; random_matched 62.2%.
• Conclusion: the hypothesis is unsupported, and the point estimate is in the opposite direction. The pattern breaks down significantly less often than every comparator: the generic low-volatility consolidation, its sideways subset, upward drift without compression, upward drift after the same impulse, and all post-impulse consolidations. Regression controls for boundary geometry reduce but do not remove the negative pattern coefficient. The full parameter grid contains no configuration with a positive lift over either low-vol baseline at any horizon.
• Chronological stability: design-period cohort A broke down at 14.1% and validation at 19.4%; the frozen out-of-sample test window (2026-07 to 2026-09) was higher at 31.1% but the generic low-vol baseline was 45.3% in the same window, so the pattern remained below it. Against the drift-without-compression baseline the test difference was +9.3% (95% CI -5.4% to +23.8%). The level is unstable across periods and no test-window comparison shows a significant elevation.
• Tradability: the short-the-breakdown strategy lost money on average at every cost level including zero (43.4% win rate, mean net return −0.40% at 0 bps/side), so even the realised breakdowns did not continue profitably.

Discipline. The primary thresholds above were fixed from the study specification before any outcome was examined. The design window (2025-01 to 2026-03) was used for diagnostics only; no threshold was selected on the validation or test windows, and the entire pre-registered grid is reported rather than its best cell. The 2026-07 to 2026-09 test window is therefore the untouched out-of-sample period.

1. Pattern definition (numeric, causal)

Timeframe 1h (resampled from 1m with the repository resampler). Window length 8 bars (variants [4, 6, 8, 10, 12]). For each bar t the window is t-w+1 .. t. All quantities use only closes/highs/lows at or before t.

• Upward drift: OLS slope of log close per bar, divided by the trailing 168h realized volatility of 1h log returns (slope_z), in [0.1, 0.6].
• Low volatility: window realized volatility / trailing 168h realized volatility (rv_ratio) ≤ 0.8.
• Prior bearish impulse: 12h log return ending one bar before the window, divided by trailing volatility × sqrt(12) (impulse_z) ≤ -1.5. The impulse is measured strictly before the consolidation.
• Frozen boundaries: pattern_high = max(high) and pattern_low = min(low) over the window; they never move after detection.
• Outcome window: starts at the open of t+1. First touch of either boundary decides. Same-hour double touches are resolved with 1m bars; same-minute touches are labelled AMBIGUOUS_SAME_BAR.

2. Data and data quality

Binance USD-M perpetual futures, 1m klines resampled to 1h, 2025-01-01 to 2026-09-30. 1m data are used for intrabar boundary ordering and the strategy test.

All symbols have complete 1m coverage over the research window (no duplicate timestamps, no missing-minute gaps), which is what makes intrabar boundary ordering resolvable. SOLUSDT carries 62,218 null cells in taker_buy_volume (unused here). HYPEUSDT starts 2025-05-30. Effective event start is later than the 200-bar warmup and the 168-bar volatility lookback. Across the primary cohorts only one same-minute ambiguity survives (cohort B); cohort A has none, so the 1m resolution removes essentially all censoring from the boundary race.

3. Event counts and primary results

Cohort A (drift + compression + prior impulse): 575 deduplicated events (one per duration-length non-overlapping window); raw qualifying bars and cluster counts are in the table. Cohort B is the same drift/compression without a qualifying prior impulse.

Cohort A by horizon

Cohort B by horizon

4. Baseline comparisons

Baselines: A random windows matched by symbol, duration, half-year, trailing-vol quintile and prior-impulse status; B generic low-volatility consolidation (no positive-drift requirement), and its sideways subset; C upward drift without volatility compression; D upward drift with the same prior impulse but no compression; E all eligible post-impulse consolidations. Pattern events are excluded from the non-random baselines to avoid self-comparison.

Pattern − baseline differences (block bootstrap, 7-day blocks)

Random-baseline match quality

Matching uses only trailing information (symbol, duration, half-year, trailing-vol quintile, prior-impulse status). Counts of match levels:

Independent-event estimate

Cohort A: 1229 raw qualifying bars, 575 deduplicated events, 564 non-overlapping clusters. All inference uses the deduplicated set; the raw count is reported so overlap remains visible.

Time to first touch and post-break dynamics (cohort A)

The post-break columns are the continuation return measured from the frozen boundary (descriptive, not a trading simulation). A positive post-breakdown value means price closed back above pattern_low; breakdowns in this sample mostly reversed (reversal within 3h 73%, 12h follow-through 31%), consistent with the negative strategy returns below.

5. Feature ablation

Cumulative conditions, 12h horizon. Adding EMA ROC/acceleration, EMA spread velocity and volume collapses the sample to single digits and is not informative.

Alternative compression definitions (drift + impulse held fixed)

Alternative prior-impulse definitions

Alternative volatility / ATR lookbacks (primary parameters)

6. Per-symbol, per-regime and per-period

Per symbol (12h)

Per regime (12h)

Chronological design / validation / test (12h)

Pattern − baseline differences by split

Resolved-event rates and 7-day block-bootstrap intervals within each chronological window. The test window is the untouched out-of-sample period.

7. Parameter sensitivity

Full predefined grid: 2700 rows (540 configurations at 12h). Configurations with a positive lift versus the generic low-vol baseline: 0; versus the sideways low-vol baseline: 0. The best 12h cell still breaks down at 33.8% versus a baseline around 51%.

8. Walk-forward validation

The frozen rows apply the pre-registered 8h definition every month. The walk_forward_selected rows choose, at the start of each month, the duration with the best lift over the low-vol baseline in all prior months and apply it to the next month. This is a genuine out-of-sample selection exercise; it is not used to choose the reported primary definition.

9. Regression controls

Linear probability model of breakdown-first on a pattern indicator plus boundary geometry (dist_high, dist_low), volatility compression, slope and impulse, with month-clustered standard errors. This asks whether pattern membership predicts breakdown beyond boundary geometry and the post-selloff environment.

{ "n": 1860.0, "coef_is_pattern": -0.22977343396220254, "ci_low": -0.2958815437355074, "ci_high": -0.16366532418889768, "p_value": 9.60568144834849e-12, "lowvol_n": 11486.0, "lowvol_coef_is_pattern": -0.08180384289025826, "lowvol_ci_low": -0.13693680668305305, "lowvol_ci_high": -0.026670879097463462, "lowvol_p_value": 0.003636233090811668 }

10. Optional strategy after costs

Pre-specified short at a confirmed breakdown below pattern_low, entry at the next 1m open, stop at pattern_high, 1.5R target, 12h time exit, costs of 0/3/6/12/20 bps per side plus realized funding. The comparison arm is a plain prior-12h-low breakdown with the same exits (simple), and one restricted to post-selloff environments (simple_impulse).

11. Limitations

• The pattern is one of many possible low-volatility drift definitions; the grid covers the pre-registered axes but not every possible normalization.
• Overlapping windows create dependent events. Counts are reported raw and deduplicated, and uncertainty uses 7-day block bootstrap, but single-symbol confidence intervals remain approximate.
• Boundary geometry is intrinsic to the pattern. The regression controls for it, but the matched random baseline matches only on the pre-registered strata.
• The chronological test window (2026-07 to 2026-09) shows a higher cohort-A rate than the design window, but it is still below the generic low-vol baseline and its difference versus the drift-only baseline has a confidence interval spanning zero. The instability is itself evidence against a stable effect.
• The optional strategy is exploratory and secondary; it is not optimized and its sample is far smaller than the event study.

Verdict: UNSUPPORTED

Primary cohort A: 95/556 resolved events broke down first within 12h (17.1%, 95% CI 14.2%–20.4%). Generic low-vol consolidation baseline broke down at 51.0%; difference -33.9% (95% CI -38.3% to -29.3%). The hypothesis is supported only if cohort A's breakdown-first rate is materially above the comparable baselines with a confidence interval that excludes zero.

Artifacts and reproduction

All machine-readable outputs are under experiments/reports/lowvol_grind_v1/: config.json, data_quality.parquet, events_cohort_a.parquet, events_cohort_b.parquet, events_baseline_*.parquet, summary_*.parquet, baseline_diffs.parquet, split_diffs.parquet, ablation.parquet, sensitivity_grid.parquet, compression_variants.parquet, impulse_variants.parquet, lookback_sweep.parquet, by_symbol.parquet, by_regime.parquet, by_split.parquet, walk_forward.parquet, regression.json, strategy_summary.parquet, summary.json and figures/.

Reproduce with:

python scripts/lowvol_grind_study.py

← All reports · source (.md)