VALTIVRA out-of-sample validation record (Methodology Specification v1.0, frozen 2026-09-04)

oos_record.csv — one row per forecast-origin month, December 1997 to August 2025 (332 months).
  origin_month          forecast origin (month-end information set)
  model_output_raw      the frozen model's one-step-ahead output at that origin, refit on point-in-time data
                        available twelve months earlier (label embargo), using exact ALFRED vintages for the
                        four revised inputs
  risk_score            percentile of model_output_raw against stored outputs dated <= origin - 12 months
                        (blank for the first 40-month warm-up, Dec 1997 - Feb 2002)
  regime                Low (<60), Elevated (60-84), High (>=85)
  recession_within_12m  1 if any NBER recession month occurred in the following twelve months, else 0

To reproduce the headline: AUROC of model_output_raw against recession_within_12m over all 332 rows = 0.82.
The 95% interval (0.67-0.94) is a 12-month block bootstrap with 1,000 draws. Regime frequency tables are
computed over the 281 rows with a risk_score. Live publication began September 2026; months after
August 2025 are not in this file because their 12-month outcome window has not closed.

This file contains model outputs only. It contains no source data and no model parameters.
Source data: FRED/ALFRED, Federal Reserve Bank of St. Louis (see the methodology page for attribution).