# Data audit
| file | SHA-256 | rows (excl. header) | cols |
|---|---|---|---|
| data/DGS10.csv | 8187f89ab6523b3c964255e1736cfcf0e2ca854ed9e50343e04f4dc2945c5cd7 (matches README) | 5217 | 2 |
| data/README.md | 32b825f02387ef3d552df6def82c585d7f248980667a28f30e58506e472e561a | provenance text | – |

- Schema: observation_date (YYYY-MM-DD), DGS10 (percent, string; empty on holidays).
- Date range 2000-01-03 … 2019-12-31; monotonic increasing; 0 duplicate dates; only weekdays (Mon 1044, Tue 1044, Wed–Fri 1043 each).
- Missing: 215 empty DGS10 fields (holidays). No non-empty unparseable values. Available values: 5002, range 1.37–6.79 %, mean 3.43 %.
- After dropping empties and differencing: 5001 changes (2500 dated ≤2009-12-31, 2501 dated ≥2010-01-01).
- 214 changes span a holiday (gap longer than the normal 1-day / Friday→Monday 3-day gap); none spans more than 4 calendar days. Kept per task rules (attributed to weekday of t).
- Information-availability: the only input is the permitted file; no violation. The forecasting weekday means use train-period changes only (asserted in code).
- Anomaly worth noting: Levene test rejects equal variances across weekdays (p≈1e-7), so the classical ANOVA's equal-variance assumption is violated; Kruskal–Wallis and Welch/HC1 checks reported as robustness.