Recession risk
Three ways to watch for a downturn
Each gauge below answers a different question and fails in a different way — that's deliberate. A near-term claims-based model, a classic 12-month yield-curve model, and a probability computed directly from GDPcast's own forecast. None of these predict an official NBER-dated recession with certainty; together they give a fuller picture than any one number.
Current readings
Last updated Oct 2, 2026, 2:25 PM ET
Near-term
Recession within 6 months
Built from initial jobless claims (level vs 24-month low, plus 3-month momentum). Beat Sahm-gap and payroll-augmented alternatives in a recursive out-of-sample horse race and has thrown no false alarm at or above 50% since 1984 — including through the entire 2022–24 inversion scare.
- Claims vs 24-month low
- -0.029
- Claims, 3-month momentum
- -0.052
Data through Oct 2, 2026
Yield curve
Recession within 12 months
Built from the 10-year minus 3-month term spread and the 6-month change in the Baa credit spread. The classic early-warning model — and, like every term-spread model including the NY Fed's, it read the 2022 inversion as high recession risk that (so far) hasn't arrived.
- 10-year minus 3-month spread
- 1.050
- Credit spread, 6-month change
- -0.310
Data through Oct 2, 2026
GDPcast-implied
Technical recession odds, from this forecast
Within 4 quarters
Within 8 quarters
Computed directly from this run's own 8-quarter GDP path and the Phase 2 backtest error distribution: the odds of two straight negative quarters (a technical recession, not necessarily an NBER-dated one) in the next 4 or 8 quarters. Moves the same morning a data release moves the forecast.
History
Near-term (6-month)
Shaded bands are NBER-dated recessions.
Yield curve (12-month)
Shaded bands are NBER-dated recessions.
These histories are fitted — computed once using today's frozen coefficients replayed over final-vintage history — not a real-time, out-of-sample record. The validation numbers below (out-of-sample AUROC, false-alarm history) are the real test of how each model would have performed live; see methodology for detail.
Methodology & validation
| Model | Estimation sample | In-sample AUROC | Recursive OOS AUROC | False alarms (≥50%) |
|---|---|---|---|---|
| Near-term | 1969–2026 (n=684) | 0.922 | 0.915 | 0 since 1984 |
| Yield curve | 1953–2025 (n=862) | 0.813 | 0.821 | 1998, Dec 2018, 2022–23, 2025 (open) |
| GDPcast-implied | Monte Carlo on this run's own path | — | — | n/a — not a fixed model, moves with each forecast |
AUROC (area under the ROC curve) measures discrimination: 0.50 is coin-flip, 1.00 is perfect separation of recession from non-recession periods. “Recursive OOS” refits each model quarterly on an expanding window starting in 1975, always using only data available before the period being predicted — the honest test of how the model would have performed live. Params version phase6c-v1.
What each gauge is actually built from
Near-term uses two transforms of weekly initial jobless claims: the level versus its 24-month low, and the 3-month momentum. Both are monthly probit inputs, chosen after a spec search against Sahm-gap and payroll-growth alternatives — those either lagged claims or threw more false alarms across 1995, 2000, 2003, 2005, 2017, and 2023–24. Unemployment itself is deliberately not an input (its coefficient goes negative once claims are included — real, but not defensible on a public page).
Yield curve uses the 10-year minus 3-month Treasury spread and the 6-monthchange in the Baa corporate bond spread over the 10-year (not the level — the level spec false-alarmed in 2012 and 2016 during ordinary credit-market wobbles).
GDPcast-implied runs a 40,000-draw Monte Carlo around the current run's own headline path, using the Phase 2 backtest's per-horizon error distribution (bias-corrected mean, bias-stripped RMSE, AR(1) correlation of ρ=0.4 across horizons) to estimate the odds of two consecutive negative quarters. A flat 2% growth path reproduces the 1984–2025 unconditional base rate almost exactly, which is the calibration check used before publishing this gauge.
Where the yield-curve gauge is known to be wrong
Every term-spread recession model, including the New York Fed's, read the 2022 yield-curve inversion as a high probability of recession that — as of this writing — has not arrived. This site's yield-curve gauge shares that miss: it peaked near 77% in mid-2023. The near-term gauge said 6% at the same time and was right. That disagreement is the reason to publish a suite instead of a single number.
Technical recession vs. NBER recession
The GDPcast-implied gauge measures a technical recession — two consecutive quarters of negative real GDP growth. The near-term and yield-curve gauges target an NBER-dated recession, the official U.S. business-cycle designation, which depends on breadth and depth across the economy, not GDP alone. The two definitions can and do diverge — 2022's two negative quarters were a technical recession but never an NBER one.
Small-sample caveat
The recursive out-of-sample tests above span six NBER recessions since 1975. Every probability here is estimated from a small number of rare events, and the maximum- likelihood standard errors on the coefficients are optimistic (the 6- and 12-month target windows overlap month to month, which serially correlates the estimation). Model choices were made on out-of-sample discrimination and false-alarm history, not on those standard errors.