Do the scores work?

Methodology

The exact rules, applied by one script.

1. Window: entries from the last 730 days of recorded scores, weekdays only.
2. Entries are sampled every 5th recorded weekday row per ticker to reduce autocorrelation between overlapping 90-day windows.
3. Forward return = price change from the entry date to 90 days later; each endpoint snaps to the nearest trading day within 7 days, and entries missing either endpoint are skipped.
4. Provenance: 11,721 entries (580 stocks) come from rows explicitly marked live or whose creation timestamp is within 7 days of their date (a historical proxy, not verified provenance); 163,327 entries (1,980 stocks) come from imported, reconciled, later-written or unknown-origin rows; the live/proxy cohort is shown separately in the Recorded live column. 871,365 rows flagged estimated are excluded, as are rows missing either score.
5. The benchmark is the equal-weight average forward return of ALL sampled entries in the same window — the covered-universe average. There is no SPY or index ticker in our price data; we do not compare against an index.
6. Median is drawn; the mean is shown with the top and bottom 1% of entries removed. Rows with fewer than 200 entries publish no statistics.
7. Win rate = share of entries with a forward return strictly above zero.
8. Every rule above is applied mechanically by one script over the raw score and price data — no hand-picking. Want to audit it? Email support — the sampling script is scripts/evidence_repro.py.