# Data audit

File: data/ip_predictors.csv — SHA-256 7e388665caf3e4e7305ae6ed8f5625ab12a6c8df5c5696c1b5dd5b86da62a39b (matches README and bundle.json).

• 417 data rows + header, 7 columns: month (YYYY-MM-01), ip_growth, oil_chg, permit_chg, neworder_chg (100×Δln, % per month), fedfunds (%, level), tcu (% of capacity, level).
• Date range 1992-01-01 … 2026-09-01; no duplicate months; no gaps (consecutive months).
• Missing: neworder_chg for 1992-01 and 1992-02 only (2 cells). No other NaN.
• Documentation discrepancy: README says the first month (1992-01) of each change series is empty, but ip_growth, oil_chg and permit_chg are populated for 1992-01 (presumably computed from 1991-12 levels). Harmless: the task's training window starts at s=1992-03 (uses ip_growth back to 1992-01 as lags; neworder_chg starts 1992-03). Recorded in data_issues.
• Outliers (|z|>5 on full sample): ip_growth 2020-04 (−14.14) and 2020-06; oil_chg 2020-03..05; permit_chg 2008-07. Genuine COVID/GFC-era observations; retained (task specifies the model exactly, no outlier handling).
• Availability: the file contains data up to 2026-09 — values dated > t exist but are only used as the realised target (2021-10…2026-09) for evaluation; the code slices data ≤ t for every origin.
• Vintage: current vintage, revised series; publication lags ignored (per README) — not real-time.

data/README.md SHA-256 225f52077afb6f435f67e2b398cbe59963ab1ce6d0d0d4e8257fd27c777ede6e.