# Data audit

• data/ip_predictors.csv — SHA-256 7e388665caf3e4e7305ae6ed8f5625ab12a6c8df5c5696c1b5dd5b86da62a39b (matches README and bundle.json). data/README.md SHA-256 225f52077afb6f435f67e2b398cbe59963ab1ce6d0d0d4e8257fd27c777ede6e.
• 417 rows × 7 columns; month 1992-01-01 … 2026-09-01, contiguous monthly, no duplicates.
• Schema: month (date), ip_growth, oil_chg, permit_chg, neworder_chg (100×Δln, percent/month), fedfunds (percent, level), tcu (percent of capacity, level). All float.
• Missing: neworder_chg 1992-01, 1992-02 only. Discrepancy with README: README says first month of every change series is empty, but ip_growth, oil_chg, permit_chg are populated for 1992-01. Used as given (only ip_growth_1992-01 enters, as lag-2 for s=1992-03).
• Anomalies: large ip_growth values 2008-09 (−4.50), 2020-03 (−3.99), 2020-04 (−14.14), 2020-06 (+6.35), 2020-07 (+3.68), 2021-02 (−3.36). Genuine events (crisis, COVID, Texas freeze); kept (spec fixed).
• Availability: file contains 2026-09 (target of last origin) — it is used only as an actual, never as an input to a forecast. Current vintage: revisions ignored per README; no real-time claim.