VALTIVRA^™

U.S. Recession Risk · point-in-time validated

Private beta. Market inputs use a research data source pending a licensed feed; recession inputs are FRED/ALFRED. Scores are published monthly and never revised.

Recession Risk Score

Trend, 3-month
12-month calibrated probability: show
Historical recession frequency in this regime

The probability is a secondary output that is not validated to beat the base rate under point-in-time data; it is hidden by default. The regime frequency is a descriptive statistic over 281 scored months, Mar 2002 – Aug 2025 — not a probability. See methodology.

Score history — walk-forward validation record (Dec 1997 – Aug 2025), then live publication

NBER recessionElevated 60–84High ≥85ScoreLive publication began

What is moving the score · inputs as of

Bars show each input's contribution to the model's log-odds of recession (right = raising risk, left = lowering it), with the input's current reading. Contributions describe the model, not the economy: strong payroll growth raises modelled risk in this specification (a late-cycle effect).

Market Stress Monitor

Composite of the inputs below, each as a percentile of its own history since 1973, equal-weighted.

Input	Reading	Percentile

The Drawdown Monitor is an analytical composite, not a statistically validated probability or forecast of future market declines. Describes current market conditions. No tested model forecast market declines reliably under point-in-time validation. Credit stress (EBP) is shown as published and is not point-in-time.

Validation record — point-in-time

0.82AUROC, 95% CI 0.67–0.94

332out-of-sample months, Dec 1997 – Aug 2025

3recessions in sample

Recession	Run-up months in High	Months in Elevated or above
Apr 2001 – Nov 2001	run-up partly outside the validation window; lead not measured
Jan 2008 – Jun 2009	12 / 12	12 / 12
Mar 2020 – Apr 2020	1 / 12	10 / 12

Regime	Months	Recession within 12 months

The regime table is computed from the published record at load time over the 281 months that carry a score (the percentile score needs a 40-month warm-up; AUROC is computed on all 332 raw outputs from Dec 1997). Same window on today's revised data scores 0.92 — the difference is the cost of data revisions in a conventional backtest, and is why the lower figure is the one reported. Live publication began September 2026; the record before that is a walk-forward simulation on archived vintages. The full series: oos_record.csv.

Known limitations

The point-in-time validation window contains three recession cycles. Performance estimates carry substantial uncertainty and no claim of consistent lead time is supportable.

2022–2024: 24 months in High with no recession following — the inverted yield curve; the same signal that gave twelve months' warning before 2008.
2005–2006: 10 months in High two years before the 2008 recession began.
2020: the model entered the Elevated-risk regime before the pandemic recession; it did not reach High. It did not predict COVID.
2017–2026: point-in-time ranking in this period is close to chance, driven by 2022–24. Three recessions in the window is a small sample.
Probabilities are not validated. No calibration construction tested achieved positive point-in-time Brier skill across periods.
Employment, production, unemployment and housing use exact ALFRED vintages; yields, sentiment and prices are treated as unrevised (unverified).