# data/README.md — F3-conf-03

## ip_predictors.csv
One row per month (`month`, YYYY-MM-01), 1992-01 to 2026-09 (417 data rows), built from six FRED series (Federal Reserve Bank of St. Louis), each retrieved with `https://fred.stlouisfed.org/graph/fredgraph.csv?id=<SERIES>` on 2026-10-30 (UTC times below). SHA-256 of this file: 7e388665caf3e4e7305ae6ed8f5625ab12a6c8df5c5696c1b5dd5b86da62a39b. License: FRED terms of use (https://fred.stlouisfed.org/legal/); the underlying series are US government data (public domain).

| column | derived from | transformation | units | retrieved (UTC) |
|---|---|---|---|---|
| `ip_growth` | INDPRO — Industrial Production: Total Index (2017=100, SA, monthly; Federal Reserve Board) | 100 × (ln INDPRO_m − ln INDPRO_{m−1}), 6 decimals | percent per month | 2026-10-30T03:20:36Z |
| `oil_chg` | MCOILWTICO — Crude Oil Prices: WTI, monthly average (USD/barrel, NSA; EIA) | 100 × Δ ln, 6 decimals | percent per month | 2026-10-30T03:24:48Z |
| `fedfunds` | FEDFUNDS — Effective Federal Funds Rate, monthly average (percent; Federal Reserve Board) | level | percent | 2026-10-30T03:24:48Z |
| `permit_chg` | PERMIT — New Privately-Owned Housing Units Authorized (thousands, SAAR; Census) | 100 × Δ ln, 6 decimals | percent per month | 2026-10-30T03:24:48Z |
| `neworder_chg` | NEWORDER — Manufacturers' New Orders: Nondefense Capital Goods Excluding Aircraft (millions USD, SA; Census) | 100 × Δ ln, 6 decimals | percent per month | 2026-10-30T03:24:48Z |
| `tcu` | TCU — Capacity Utilization: Total Index (percent of capacity, SA; Federal Reserve Board) | level | percent | 2026-10-30T03:24:48Z |

Empty cells: the first month of each change series (1992-01) and `neworder_chg` for 1992-01/02 (NEWORDER starts 1992-02). No other missing values.

**Vintage caveat**: all series are the current (final) vintage as of retrieval; INDPRO, TCU and NEWORDER are revised substantially after first publication, and publication lags are ignored in this exercise (a value dated month m is treated as known at the end of month m). Results therefore measure the predictability of the revised series under a stylised timing assumption — not historical real-time performance.