50-name mega-cap panel | abnormal log views vs 28d median, lag-1 | 2018-10-31..2026-08-05 (1949 days) | Bonferroni bar p<0.00833 over 6.
Cross-sectional Spearman IC (per-date permutation null x200)
| horizon | mean IC | partial IC|mom5 | p(cont) | p(rev) | n days | |---|---|---|---|---|---| | 1d | +0.0049 | -0.0033 | 0.060 | 0.940 | 1949 | | 5d | +0.0044 | +0.0042 | 0.110 | 0.890 | 1949 | | 21d | -0.0047 | -0.0035 | 0.945 | 0.055 | 1933 |
Equal-weight panel buy&hold: Sharpe 1.05, ret +322.5%, DD -32.5%
Tradable top/bottom-decile (next-open fills)
| hold(d) | bps | leg | Sharpe | ret | maxDD |
|---|---|---|---|---|---|
| 1 | 10 | top | 0.22 | +20.7% | -34.0% |
| 1 | 10 | bottom | -0.15 | -37.5% | -48.3% |
| 1 | 25 | top | -1.27 | -90.3% | -91.0% |
| 1 | 25 | bottom | -1.62 | -95.4% | -95.5% |
| 5 | 10 | top | 0.77 | +199.6% | -36.7% |
| 5 | 25 | top | 0.31 | +39.1% | -37.4% |
Verdict
REJECT — no attention leg clears the Bonferroni permutation bar in the pre-registered direction; tradable deciles must also beat equal-weight net.
Reading
- Continuation = buy high-attention names; reversal = buy low-attention (Barber-Odean attention pop then fade).
- Partial IC controls the mode-2 leakage: attention is partly caused by past price moves, so a raw IC that dies on partial is momentum. Runtime 5s.