Fishing scan with a block-permutation null (2000 perms) and a Bonferroni bar across the 7-series panel (p < 0.0071). Signal = 252d z of the series, lagged 1 day (PIT). Target = SPY log return that day. A finding is only interesting if p < Bonferroni bar.
| topic | n | corr(next-day ret) | perm p | survives |
|---|---|---|---|---|
| space_weather_Ap | 5698 | +0.0013 | 0.9245 | no |
| sunspot_number | 5698 | +0.0081 | 0.5340 | no |
| RRPONTSYD | 3357 | +0.0069 | 0.6815 | no |
| SOFR | 2053 | -0.0208 | 0.3370 | no |
| DFF | 5576 | -0.0000 | 0.9990 | no |
| DTWEXBGS | 5138 | -0.0047 | 0.7480 | no |
| DCOILWTICO | 5696 | -0.0100 | 0.4560 | no |
Reading
- Any
survives=yesrow is a candidate for a real study (IS/OOS, cost ladder, placebo). Absent that, the topic is noise on this sample — same lesson asreports/placebo_audit.md. - Correlations here are small by construction; the null, not the sign, is the point.