# Data audit (P1)

File: data/canada_gdp_employment.csv — SHA-256 f4c8e4976fb5f5f0c326e40bcb4f85b8b22963e5be89e2729e1a430c198f2bed (matches README and bundle.json). data/README.md SHA-256 2000e2e6...fa2f (matches bundle).

• 345 data rows x 5 columns (month, gdp_level, gdp_yoy, emp_level, emp_yoy). Months 1998-01-01..2026-09-01, contiguous monthly, no duplicate months.
• Units: gdp_level chained 2017 $ millions SAAR; emp_level thousands SA; *_yoy percent (4 dp).
• Missing: gdp_level/gdp_yoy missing only for 2026-08, 2026-09 (not yet published, expected). No other missing values.
• Consistency: gdp_yoy and emp_yoy recomputed from levels (100*(L_m/L_{m-12}-1), rounded 4dp) match exactly for all months from 1999-01 (max abs diff 0.0).
• Anomaly: README says yoy is "empty for the first 12 months", but gdp_yoy and emp_yoy are populated for 1998-01..1998-12 (presumably computed from 1997 levels not in the file). Cannot be verified from file; values used as given (only affect early training samples).
• Anomaly: COVID-19 period: gdp_yoy ranges -16.2 (2020-04) to +18.5 (2021-04); emp_yoy -15.6 to +16.1. Extreme outliers relative to the rest of the sample (sd ~2.7). Handling: training observations whose target or regressor months fall in 2020-03..2021-06 are excluded from model estimation (pre-specified, not tuned on test errors); naive unaffected; all test targets (2022-06..2026-07) lie outside this window.
• Availability: file contains data beyond each origin (current vintage up to 2026-09). Handling: analysis.py slices per origin (GDP <= t-2, emp <= t-1) and records last dates used per forecast; checks assert the rule.
• Vintage: final-vintage data, so exercise is pseudo-real-time (timing only).