September 11, 2026 · S&P 500 at 7,592 (2026-09-10) · daily closes (Yahoo Finance)
What this actually predicts
Of the last 21 ≥10% corrections the model had the history to classify, 21 of 21 started while it read calm (at or below the base rate), and 21 began above the 200-day trend.
That is almost by construction — a correction starts at a high, and at a high this gauge is calm — which is exactly the point: this is a continuation-and-severity gauge, not an onset predictor. It's good at telling you a decline already underway is likely to deepen (below trend, choppy). It cannot call the top, and a calm reading today is the same reading that preceded 2000, 2007, COVID and 2022.
The gauges
Trend+6.1% vs 200-day averageGREEN
Big drops overwhelmingly start in markets already below their long average.
Volatility21st percentile (21-day realized)GREEN
Turbulence clusters — choppy markets fall further. Red above the 85th percentile.
Drawdown-2.7% from recent highGREEN
Drops cluster — already being 5%+ off the high has meant ~1.6x the odds of extending to -10%.
Credit stress (HY spreads)2.71%, +0.01pp this monthGREEN
Credit usually cracks before stocks; widening of 0.5pp+ in a month is the flag. One caveat this cycle: private credit can hide stress that used to show up here.
Financial conditions (NFCI)-0.56 (loose)GREEN
Chicago Fed's broad weekly index — the strongest published 1-month warning. Positive = tighter than average.
Layoffs (jobless claims)206k/week, +4% above the year's lowGREEN
New unemployment filings, 4-week average. Rising layoffs mark the slow, recession-type declines — the ones that go deepest. Warning at +10% off the low, red at +20%.
Safe-company spreads (IG)0.81% — quietGREEN
What solid, investment-grade companies pay to borrow above Treasuries. Junk spreads widening alone means trouble in risky corners; both rising together means trouble everywhere.
VIX term structure0.87 (contango)GREEN
Near-term fear above 3-month fear (ratio over 1) has marked every major selloff — but it confirms one underway rather than predicting it. Red only when sustained 5+ days.
Context — shapes how bad a drop gets, not when it starts (not in the odds)
Yield curve (10yr minus 3mo)+0.95% — recently un-inverted — historically the risky windowYELLOW
The bond market's recession warning, 6-18 months ahead — far too slow to time a correction, but it tells you whether a decline would likely be the deep, recession-driven kind.
Sahm rule-0.07pp (as of 2026-08)GREEN
Recession marker (triggers at +0.50). Doesn't time corrections, but recessionary declines run deeper and longer than non-recessionary ones.
Valuation (Shiller CAPE)40.7 (99th percentile, live)RED
Doesn't time anything at 1-3 months — it changes how far drops go once they start. Rich valuations amplify the same shock. Live from multpl.com; percentile vs the full 1871-present Shiller history.
How the odds are computed (and the full history table)
Two questions sort every trading day since 1950 into a box: is the market above or below its long average (trend), and how choppy has it been (volatility, ranked against a trailing ~10-year window so no future information leaks in)? The number is how often the close then fell ≥10% from that day within the horizon. No model, no weights — just counting.
On the sample size: the pill counts spells and events, not days — because 2,000 consecutive days in one regime aren't 2,000 independent trials, they're a few dozen episodes sharing overlapping forward windows and the same handful of corrections. That's why the intervals are wide and several cells overlap. The ordering (below-trend/choppy ≫ above-trend/calm) is the robust part, and it holds whether you set the bar at 5%, 10%, 15% or 20%.
3 months (average across all days: 11.0%)
below trend, high vol24.8% [14–32%]119 spells · 63 events
below trend, mid vol29.1% [21–36%]227 spells · 79 events
below trend, low vol12.0% [7–19%]138 spells · 27 events
above trend, high vol9.1% [5–13%]124 spells · 28 events
above trend, mid vol9.0% [6–12%]401 spells · 66 events
above trend, low vol ← today6.0% [4–8%]322 spells · 61 events
The bracket is a block-bootstrap 90% interval resampled over regime spells, not days — the honest uncertainty. Note how widely the cells overlap: the real resolution is “below trend & choppy” vs “above trend & calm,” not six distinct numbers.
1 month (average across all days: 3.0%)
below trend, high vol10.9% [4–16%]119 spells · 40 events
below trend, mid vol9.0% [5–12%]227 spells · 31 events
below trend, low vol1.2% [0–3%]138 spells · 6 events
above trend, high vol4.0% [1–7%]124 spells · 11 events
above trend, mid vol2.0% [1–3%]402 spells · 23 events
above trend, low vol ← today0.8% [0–1%]323 spells · 19 events
The bracket is a block-bootstrap 90% interval resampled over regime spells, not days — the honest uncertainty. Note how widely the cells overlap: the real resolution is “below trend & choppy” vs “above trend & calm,” not six distinct numbers.
Read this before acting
This catches slow-building declines — credit widening, trend breaks, and volatility regime shifts that unfold over weeks (2000, 2008, 2022). It cannot see news shocks: COVID read 1-in-20 the day before a 34% crash, and October 1987, August 2024, and April 2025 looked just as calm. Against shocks, only standing hedges protect you — never signals. A low reading means no internal fragility, not no risk.
One known unknown this cycle: private credit can hide stress that used to show up in the high-yield spread, so a quiet credit gauge is weaker evidence of calm than it once was. The probabilities here are measured frequencies — trust the cells built on many events.