Vol × Session (Asia/EU/US): broader bucket sibling of vol×hour
DroppedVol × Session (Asia/EU/US): broader bucket sibling of vol×hour
2026-05-17 · status: dropped · 0.9s
Hypothesis: At least one (vol_regime × session × weekend) cell produces systematically non-zero forward 4h returns walk-forward — sample size per cell roughly 8× larger than the per-hour analysis, which may surface a signal that 24-hour slicing averaged into noise.
Verdict: DEAD — strongest cell only +10.2 bps (t +1.69). Session aggregation didn't recover a signal — confirms the vol×time class is honestly closed.
Key metrics
| metric | value |
|---|---|
| n_cells | 9 |
| n_robust | 0 |
| strongest_cell | vol_Q3 × US × weekday |
| strongest_mean_bps | +10.2087 |
| strongest_t_stat | +1.6850 |
| strongest_stability | +0.6667 |
| n_windows | 21 |
Approach
3 sessions (Asia 0-8 UTC, EU 8-16 UTC, US 16-24 UTC) × 3 vol regimes × 2 weekday/weekend flag = 18 cells. One observation per session per day (anchor at mid-session hour 4/12/20 UTC) → within-cell obs are exactly 24h apart, independent samples.
Promote gate: ≥ 1 cell with pooled |mean fwd-4h| > 20 bps, |t-stat| > 2.0, stability ≥ 70% across walk-forward windows that have ≥ 30 obs in the cell.
All cells (sorted by |mean|)
| cell | n_total | n_windows | mean_bps | se_bps | t_stat | stability |
|---|---|---|---|---|---|---|
| vol_Q3 × US × weekday | 606 | 12 | 10.2 | 6.06 | 1.69 | 0.67 |
| vol_Q3 × Asia × weekday | 293 | 7 | 9.7 | 8.83 | 1.1 | 0.57 |
| vol_Q1 × EU × weekday | 457 | 11 | -8.2 | 6.13 | -1.34 | 0.64 |
| vol_Q2 × EU × weekday | 187 | 5 | -8.2 | 11.57 | -0.71 | 0.8 |
| vol_Q2 × US × weekday | 105 | 3 | 6.6 | 9.49 | 0.69 | 0.67 |
| vol_Q1 × US × weekday | 39 | 1 | 5.9 | 8.59 | 0.69 | 1 |
| vol_Q2 × Asia × weekday | 142 | 4 | 4 | 7.15 | 0.56 | 0.75 |
| vol_Q1 × Asia × weekday | 248 | 6 | 3 | 4.03 | 0.75 | 0.67 |
| vol_Q3 × EU × weekday | 176 | 4 | 1.3 | 18.4 | 0.07 | 0.75 |
Robust cells
None — no cell passes all three gates.
