# Data audit

Schema: month (date, YYYY-MM-01, 1998-01..2026-09, consecutive monthly, no duplicates); gdp_level (float, millions chained 2017 $, SAAR); gdp_yoy (percent); emp_level (float, thousands); emp_yoy (percent).

Missing: gdp_level/gdp_yoy missing for 2026-08, 2026-09 only (not yet published, as documented). emp complete.

Consistency: recomputed 100*(x_m/x_{m-12}-1) matches the given yoy columns to ≤5e-5 (rounding) for all months 1999-01 onward, both series.

Anomalies:

1. README says gdp_yoy is "empty for the first 12 months", but 1998 rows have gdp_yoy and emp_yoy values. They cannot be verified from levels in the file (no 1997 levels). Handling: excluded from estimation samples (logged in exclusions.csv); they never enter as targets anyway for the test period.
2. COVID-19 period (2020-03..2021-12): extreme y/y values (GDP −16.2% to +18.5%) from the lockdown and the subsequent base effects. Handling (pre-specified, known at the first origin 2022-05): training pairs whose target or regressor month falls in 2020-03..2021-12 are dropped from model estimation. A sensitivity run without this exclusion is reported.

Information availability: the file contains future data relative to each origin (current vintage, all months). Handling: analysis.py truncates per origin to GDP ≤ t−2 and employment ≤ t−1 before any estimation, and records last dates used per forecast; vintage (revision) issue cannot be fixed — no real-time claim is made.