# Data audit

• File: data/ip_predictors.csv, SHA-256 7e388665caf3e4e7305ae6ed8f5625ab12a6c8df5c5696c1b5dd5b86da62a39b (matches README and bundle.json).
• 417 data rows × 7 columns (month, ip_growth, oil_chg, fedfunds, permit_chg, neworder_chg, tcu); all numeric float except month (YYYY-MM-01). Range 1992-01-01 … 2026-09-01, consecutive months, no duplicates.
• Units: ip_growth, oil_chg, permit_chg, neworder_chg = 100×Δln (percent/month); fedfunds percent level; tcu percent of capacity.
• Missing: only neworder_chg for 1992-01 and 1992-02 (2 cells). These months are never used as ridge feature rows (training starts s=1992-03) so no exclusion needed.
• Anomaly/discrepancy: README says the first month (1992-01) of every change series is empty, but ip_growth, oil_chg and permit_chg have values in 1992-01. Used as given (ip_growth 1992-01 is needed as lag-2 for s=1992-03). Recorded in data_issues.
• Outliers (|z|>5, full sample, for audit only): ip_growth 2020-04 (−14.14), 2020-06 (+6.35); oil_chg 2020-03/04/05 (±55); permit_chg 2008-07 (−24.8). Genuine COVID/GFC-era movements; kept (task specifies no outlier treatment). They lie in the training windows, not in the evaluation period.
• Availability: file contains 2026-09 (dated > last origin 2026-08); it is used only as the target of the last origin. Code slices data to ≤ t for every fitting step.
• Vintage: current vintage; publication lags ignored per README — not real-time.