Data audit
| file | SHA-256 | rows (excl. header) | cols |
|---|---|---|---|
| data/gc5y_daily.csv | 2cdda6ccb2147c8861997166cb0d32cf85ff2744e1e7c0f47eded0f354c2db51 (matches README) | 2498 | 3 |
| data/README.md | 7716f8b58d8f4729504a7713a8a3d651cd2f7298f53f668b377215a5b2af1e34 | – | – |
- Schema: date (YYYY-MM-DD string), yield_5y (percent, 2 decimals), post_cut (int).
- Date range: 2010-01-04 to 2019-12-31; sorted ascending; 0 duplicate dates; 0 missing values.
- yield_5y range 0.48–3.20 %, mean 1.574.
- Calendar gaps between rows: 1 day (1962), 2 (10), 3 (440, weekends), 4 (79), 5 (6, year-end holidays). Plausible business-day calendar.
- post_cut: all 2498 rows = 0. Consistent with the date rule (no dates ≥ 2020-03-04). README confirms retrieval used end_date=2019-12-31.
- Critical anomaly: the treatment window is entirely absent, so the estimand (post minus 120-day pre window anchored on the first flagged day) cannot be computed. Decision: declare premise_invalid / infeasible; do not use proxies (e.g., last 120 days of 2019 vs nothing) or external data.
- No input violates the information-availability rules.