503-name panel. Signal = past-k overnight return (clipped +/-20%), lag-1; target = next-day close-to-close. IC = per-date Spearman. Reversal expects IC<0. Turnover-free, so no cost/turnover confound.
| signal | full IC | t | IS IC | OOS IC | Q5-Q1 spread (full) |
|---|---|---|---|---|---|
| overnight k=1 | -0.0041 | -1.69 | -0.0066 | -0.0018 | +0.0002 |
| overnight k=5 | -0.0061 | -2.90 | -0.0070 | -0.0053 | +0.0004 |
| overnight k=21 | -0.0013 | -0.71 | -0.0030 | +0.0002 | +0.0000 |
| intraday k=5 | +0.0063 | +2.65 | -0.0003 | +0.0122 | -0.0004 |
Placebo (per-date permutation, 100 draws) on overnight k=5: observed IC -0.0061, null sd 0.0007, two-sided p = 0.000.
Verdict
REJECT — cross-sectional overnight information too small: |IC| 0.0061 (need >= 0.01 and p<0.05), p 0.000. The reversal sign is consistent IS/OOS, so the ranking is real but the quintile spread (~4bps/day) is far below the turnover cost of a daily long-short book — statistically present, economically not tradeable as a sleeve.
Time 219s.