Lab · ML Experiments

ML — Pattern Discovery

Inverted workflow: find conditional edges in BTC data first, build strategies second.
55 experiments

Vol-regime transitions: directional reaction at boundary crossings

Dropped
2026-05-17 volatilityregimetransitionsevents
Hypothesis
When trailing rv_4h transitions between regime terciles (low/mid/high), BTC shows a directional bias in the next 4h that's distinct from non-transition hours. At least one (from→to) cell passes |mean| > 15 bps, |t| > 2, stability ≥ 60% walk-forward.
Verdict
**DEAD** — strongest transition cell only **+6.2 bps** (t **+0.14**). Vol regime *changes* are no more informative than vol *levels* — direction at the transition is noise. Confirms vol-as-sizing is the only edge in this data.
n_cells
6
n_robust
0
n_windows
21
strongest_cell
HIGH → LOW
strongest_t_stat
+0.1387
strongest_mean_bps
+6.1792
n_transition_events
3,757
strongest_stability
+1.0000

Vol-regime transitions: directional reaction at boundary crossings

2026-05-17 · status: dropped · 0.7s

Hypothesis: When trailing rv_4h transitions between regime terciles (low/mid/high), BTC shows a directional bias in the next 4h that's distinct from non-transition hours. At least one (from→to) cell passes |mean| > 15 bps, |t| > 2, stability ≥ 60% walk-forward.

Verdict: DEAD — strongest transition cell only +6.2 bps (t +0.14). Vol regime changes are no more informative than vol levels — direction at the transition is noise. Confirms vol-as-sizing is the only edge in this data.

Key metrics

metric value
n_transition_events 3,757
n_cells 6
n_robust 0
strongest_cell HIGH → LOW
strongest_mean_bps +6.1792
strongest_t_stat +0.1387
strongest_stability +1.0000
n_windows 21

Approach

Hourly rv_4h classified into 3 terciles using cuts learned on each walk-forward training segment, applied to test. A transition is hour t where the regime differs from hour t-1 AND the previous regime held for ≥ 4h (filters out single-bar noise). Forward 4h return computed at transition timestamp.

6 transition cells (3×3 minus diagonal). Promote gate: ≥ 1 cell with pooled |mean| > 15 bps, |t| > 2.0, stability ≥ 60% across walk-forward windows.

Total sustained transition events: 3,757

Per-transition results

cell n_total n_windows mean_bps se_bps t_stat stability
HIGH → LOW 5 1 6.2 44.55 0.14 1
LOW → MID 969 20 3 3.99 0.76 0.65
MID → HIGH 607 21 2.5 6.15 0.41 0.48
LOW → HIGH 70 9 -2.3 20.41 -0.11 0.44
MID → LOW 896 20 -1.2 3.39 -0.36 0.45
HIGH → MID 1169 21 0 3.32 0 0.43

Robust cells

None — no transition cell passes all three gates.

transition heatmap