Data audit
| file | SHA-256 | data rows | cols | dates | missing cpi | dup dates |
|---|---|---|---|---|---|---|
| cpi_vintage_2016-07-15.csv | 281fbcdd6a6eeb3e9cb7c1f20476e6bd1f7b925bf140606fe46538fe83f2c83e | 834 | date,cpi | 1947-01..2016-06, no gaps | 0 | 0 |
| cpi_vintage_2026-10-28.csv | 8220d5b4d70ca32fc2c7b1975651c515967db8a87053cf725159c63ed274cfba | 153 | date,cpi | 2014-01..2026-09, no gaps | 1 (2025-10) | 0 |
| cpi_assembled.csv | ff732b201a525d6fb171d4e8b2d470f986863c1547a1928e42bb197c0777d44d | 957 | date,cpi,vintage | 1947-01..2026-10, gap at 2016-07 | 1 (labelled 2025-11) | 0 |
Note: README.md lists SHA-256 values "of the raw download" which differ from the file hashes above; README says the downloads were reformatted, so a mismatch is expected and not itself evidence of corruption.
Units: index 1982–84=100. Range 21.48–336.298; no month-on-month jump >3%.
Anomalies:
- 2026 vintage: empty cpi at 2025-10 (source reports no value). Not imputed.
- Overlap 2014-01..2016-06 (30 months): all 30 values differ between vintages (new−old in [−0.455, 0.330]). Rule: use 2026 vintage.
- Assembled file (hand-made, not authoritative): (a) uses 2016 vintage for 2014-01..2016-06, violating the rule; (b) 2016-07 missing and all 123 subsequent rows shifted one month forward (value at label t = vintage value of t−1), producing a spurious 2026-10 row and placing the missing value at 2025-11. Decision: the assembled file is read only for audit and is NOT used for computation.
- Mixed-vintage growth: g_t for 2014-01..2014-12 uses P_t (2026 vintage) and P_{t−12} (2016 vintage), as the rule implies. Kept; sensitivity for 2014-06 reported.