← Stage Two ResearchUS Recession Indicator
September 11, 2026
Cycle state
late-cycle deterioration
The forward bucket is elevated (curve inverted or in the post-inversion danger window / credit widening) while the nowcast is still healthy. This is what the start of every cycle-end looks like — and it's exactly the signal a single averaged score buries.
Nowcast — are we in one now?0/100LOW
Expansion-consistent readings.
Forward — next 12 months23/100MODERATE
Some deterioration — watch closely.
Overall9/100LOW
Expansion-consistent readings.
What the forward score means: the bands aren't arbitrary. Since 1960, when the forward score sat in the MODERATE band, an NBER recession began within 12 months ~14% of the time [8–22%], vs a 13% base rate for any random 12-month window. That interval rests on only 5 distinct recessions, so it's wide on purpose — with seven-ish recessions since 1970, no recession model can be more precise than this, and any that claims to be is overfit.
Nowcast indicators
Coincident measures. The Sahm rule dominates by design — it has marked every recession since 1970 with essentially no real-time false positives.
Sahm rule-0.07ppGREEN
3-mo avg unemployment 4.13% vs 12-mo low 4.20%. Triggers at +0.50pp, warning at +0.30pp. weight 25
Initial claims206k (+4%)GREEN
4-week average vs 52-week low of 199k. Warning at +10%, red at +20%. weight 15
Payroll growth+71k/moGREEN
3-month average monthly change. Warning below +50k, red below zero. weight 12
Industrial production+1.1% YoYGREEN
Warning below 0%, red below -1%. Can false-positive in manufacturing-only slowdowns. weight 9
Real retail sales+2.5% YoYGREEN
Warning below 0%, red below -2%. weight 7
Indeed job postings-1.1% YoYGREEN
Daily labor-demand proxy. Warning below -10%, red below -20% (loose on purpose — prone to false positives). weight 5
Forward indicators
Leading measures, 6–18 months. The curve gives direction, not timing. The old level-based curve probit was dropped — it double-counted the same 10y-3m spread and read “safe” exactly as the curve re-steepened into 2001, 2008 and 2020 — and replaced with a post-inversion clock that scores the window where recession odds actually stay high.
Yield curve (10y-3m)+0.95%YELLOW
re-steepened after inversion (danger window) — ~22% went on to a recession within 12 months, vs 5% when the curve never inverted; 12-month low -0.07%. weight 15
Post-inversion clock12mo since un-inversionYELLOW
Months since the 10y-3m curve exited a sustained inversion — inside the post-inversion danger window (recession odds stay ~40%+ for roughly a year after un-inversion). Replaces a level-based probit that goes falsely calm right as the curve re-steepens into a recession. weight 10
High-yield credit spread2.71% (13th pct)GREEN
Ranked against its own trailing ~10-year range (warning 70th percentile, red 90th) — the index has drifted up in quality, so fixed levels mislead. One blind spot to carry from the correction page: private credit can now hide stress that used to show up here. weight 8
Building permits+2.4% YoYGREEN
Warning below -5%, red below -15%. Permits lead starts. weight 7
VIX (21-day avg)15.2GREEN
Warning at 25+, red at 35+. Sustained, not single-day, levels. weight 5
Benchmark · Chauvet-Piger model0.8% as of 2026-07-01
Published Markov-switching benchmark. Lags ~3 months; >80% historically confirms recession. Not in the composite.
Can this be backtested? (read this before acting)
Honestly, not in the usual sense. There have been ~7 US recessions since 1970. Every recession model ever built — Sahm, Estrella, Chauvet-Piger, the Conference Board LEI — is fit to single-digit effective N. IC math doesn't apply and walk-forward barely does. So instead of pretending, this page does two honest things: it counts (the forward bands above show the actual historical recession-within-12-months rate, with deliberately wide intervals), and it leans on components with published real-time records.
The Sahm rule has an unmatched real-time nowcast record; initial claims and payroll momentum are the standard coincident core; the yield curve and credit spreads are the canonical forward stack. The composite inherits whatever validity those pieces have — no more, no less. It cannot see shocks (a COVID-style hit moves everything in weeks), and it shares the correction page's blind spot: private credit can now hide funding stress that used to show up in the HY spread.
Read the cycle state and the two bucket scores, not the single overall number. The most useful configuration this page can flag is “forward elevated, nowcast still clean” — late-cycle deterioration — which is where we are now: past a curve inversion, in the danger window, with coincident data still healthy.
Score bands: under 15 low · 15–35 moderate · 35–60 elevated · 60+ high, anchored to counted recession frequencies (above). Each indicator scores on a piecewise-linear ramp between its green and red thresholds (not a 0/0.5/1 step, so one soft print can't flip the score overnight), weighted and normalized per bucket. Thresholds and weights are judgment calls, chosen for transparency over fitted precision at this sample size.