# Data audit

Schema: month (YYYY-MM-01, 1992-01-01..2026-09-01, 417 consecutive months, no gaps, no duplicates); ip_growth, oil_chg, permit_chg, neworder_chg (percent per month, 100×Δln); fedfunds (percent, level); tcu (percent of capacity, level). All float64.

Missing values: only neworder_chg at 1992-01 and 1992-02 (2 cells). README states that the first month (1992-01) of every change series is empty, but ip_growth, oil_chg and permit_chg are populated for 1992-01 — a minor README/data inconsistency. It has no effect: the earliest value used is ip_growth for 1992-01 (as lag 2 of training row 1992-03), which is present; neworder_chg is first needed at 1992-03, present.

Ranges (min / max): ip_growth −14.14 / 6.35; oil_chg −56.8 / 54.6; fedfunds 0.05 / 6.54; permit_chg −24.8 / 17.1; neworder_chg −11.4 / 9.1; tcu 64.1 / 85.0. Magnitudes plausible for units.

Anomalies: |z|>5 outliers — ip_growth 2020-04 (−14.14) and 2020-06 (+6.35); oil_chg 2020-03/04/05 (COVID); permit_chg 2008-07. These are genuine economic events; retained (ridge spec fixes training rows; no robustification allowed by the spec). They lie in every training window and influence μ, σ, b.

Information availability: the file contains data up to 2026-09, which is the target of the last origin (2026-08). The script masks all rows dated > t at every origin before computing any statistic. No input violates the rules per se; the vintage (current, revised) and ignored publication lags are accepted stylisations stated in TASK.md/README.