# Data audit
| file | SHA-256 | rows | cols |
|---|---|---|---|
| data/ip_predictors.csv | 7e388665caf3e4e7305ae6ed8f5625ab12a6c8df5c5696c1b5dd5b86da62a39b (matches README & bundle.json) | 417 | 7 |
| data/README.md | 225f52077afb6f435f67e2b398cbe59963ab1ce6d0d0d4e8257fd27c777ede6e | – | – |

Schema: month (YYYY-MM-01, 1992-01-01..2026-09-01, contiguous monthly, 0 duplicates); ip_growth, oil_chg, permit_chg, neworder_chg (percent/month, 100×Δln); fedfunds (percent level); tcu (percent of capacity). All float.

Missing: neworder_chg 1992-01 and 1992-02 only (2 cells). No other missing.

Anomaly / 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 in 1992-01. Handling: values used as given; only ip_growth_1992-01 enters the analysis (as the lag-2 feature of the first training row s=1992-03). No impact on the specified windows otherwise.

Extremes (|z|>5 full sample, descriptive only): ip_growth 2020-04 (−14.14), 2020-06; oil_chg 2020-03..05; permit_chg 2008-07. Genuine COVID/GFC movements; kept (task prescribes the sample, no outlier rule).

Availability: file contains data through 2026-09 (later than most origins); the script slices data ≤ t per origin; 2026-09 is used only as the scoring target of the last origin. Current vintage, publication lags ignored (stated in README) — reported as limitation.