Vol × Time interaction: hidden cells in (regime × hour × dow)

2026-05-17 · status: inconclusive · 1.5s

Hypothesis: Certain (vol_regime × hour) or (vol_regime × day-of-week) cells produce systematically non-zero forward 4h returns walk-forward, even though the aggregate hour/dow effects were near-zero. Cells with within-cell sample independence (24h apart) give clean t-stats.

Verdict: MARGINAL — strongest cell vol_Q3 × hour=07 shows mean +18.6 bps, t +1.73, stability 75% — close to the bar but not over. Tighter horizons (1h/15m) or finer regime bins might surface a clearer effect, but for now no cell meets the production gate.

Key metrics

Background — what we've already ruled out

Calendar/time factors we've previously tested in isolation and found weak or dead:

This experiment: cross the two surviving findings (vol-clustering & weak time effects) into a 2D grid. Maybe opposing effects in different vol regimes cancel each other in aggregate analysis.

Approach

For each walk-forward window: vol-regime terciles are learned on the training segment (cuts based on trailing rv_240m), then applied to the test segment. Each test-set hourly observation belongs to one of 3 vol regimes and exactly one hour of day / day of week. Within a single (regime × hour) cell, observations are spaced 24h apart → independent samples → t-stats are honest.

Cells: 3 × 24 = 72 (vol × hour), 3 × 7 = 21 (vol × dow), 3 × 24 × 7 = 504 (full 3D, only checked if 2D passes).

Promote gate — at least ONE cell with: - pooled |mean fwd-4h| > 20 bps - |t-stat| > 2.0 - same sign in ≥ 70% of walk-forward windows that have ≥ 10 obs in the cell

1. Vol-regime × Hour-of-day

Mean fwd-4h return (bps) per cell

t-stat per cell (|t| > 2 ≈ unlikely zero)

Stability (fraction of WF windows agreeing with overall sign)

2. Vol-regime × Day-of-week

Mean fwd-4h return (bps) per cell

t-stat per cell

Stability

3. Robust cells (passing all gates)

No cells pass all three gates.