# Data audit

# data/ip_predictors.csv

SHA-256: 7e388665caf3e4e7305ae6ed8f5625ab12a6c8df5c5696c1b5dd5b86da62a39b (matches data/README.md and bundle.json)
417 data rows × 7 columns (month, ip_growth, oil_chg, fedfunds, permit_chg, neworder_chg, tcu); all numeric float64 except month string.
Range 1992-01-01 .. 2026-09-01; consecutive months, no gaps, no duplicate months.
Units: change columns = 100·Δln (percent/month); fedfunds percent level; tcu percent of capacity.
Missing: only neworder_chg at 1992-01 and 1992-02 (2 cells).
Anomaly vs README: README says the first month (1992-01) of every change series is empty, but ip_growth, oil_chg and permit_chg have values at 1992-01 (presumably computed from 1991-12 levels). Handling: values used as given; only ip_growth 1992-01 enters the analysis (as lag-2 feature of the s=1992-03 training row, as the task specification requires). Recorded in data_issues.
Extremes (|z|>5 full-sample, reported not removed): ip_growth 2020-04 (−14.14) and 2020-06 (+6.35); oil_chg 2020-03..05; permit_chg 2008-07. These are genuine COVID/2008 episodes; kept (task specifies the model exactly, no outlier treatment).
data/README.md: SHA-256 225f5207...ede6e (documentation only).

No column is dated after its month label; publication lags are ignored by task design. Evaluation targets 2021-10..2026-09 all present. Rows 1992-01/02 are not training rows (training starts s=1992-03, whose features need ip_growth 1992-01..03 and neworder_chg 1992-03 — all present).
Vintage: current vintage (revised data) — results not real-time; stated in report.