Before the Crash, the "Crash Index" Was Asleep

The CBOE SKEW index is marketed as the market's tail-risk gauge, the black swan index. It is built from the prices of out-of-the-money S&P 500 options, and the pitch is intuitive: when traders pay up for crash protection, SKEW rises, and a rising SKEW is supposed to mean a crash is more likely. We pulled all 36 years of it, every trading day from 1990 to today, and asked the only question that matters. When SKEW was high, did the S&P actually fall more often? It did not. If anything, the opposite. It is the second Cboe index sold as a crash signal that we have put on the stand; the first, the dispersion index, failed the same way.

High SKEW, same market

Start with the simplest test. Sort every day by how high SKEW was, then look at what the S&P did next. If SKEW were a crash gauge, the high readings should be followed by weaker returns and more drawdowns. They are followed by neither.

Forward S&P 500 by SKEW percentile, 1990 to 2026. As SKEW rises, returns hold up, the share of positive months rises, and the odds of a crash fall. The opposite of what a tail-risk gauge should do.

SKEW reading	Next month (median)	Positive	Odds of a 10% drop (30d)
Lowest half (under 50th pctile)	+1.2%	63%	7%
50th to 80th	+1.2%	63%	7%
80th to 90th	+1.3%	66%	5%
90th to 95th	+1.2%	68%	4%
Top 5% (highest SKEW)	+1.6%	69%	3%
Any day (baseline)	+1.2%	64%	7%

The days when SKEW was screaming the loudest about tail risk were followed by the fewest crashes, a 3% chance of a 10% drop against a 7% baseline, and slightly more up months than average. Absolute levels say the same: whether SKEW sat at 120 or 145, the next sixty days returned about +3% and carried roughly the same one-in-ten chance of a 10% drawdown. No level of SKEW separated the dangerous days from the calm ones. The correlation between SKEW and the next month's return is -0.02; between SKEW and the next sixty days' worst drawdown, +0.02. Both are zero. A crash gauge would show a clear negative number.

It was not there before the crashes

Forget thresholds and go straight to the crashes. We found every market peak that was followed by a drop of more than 10% within two months, and looked up where SKEW was sitting at the time. The median across all sixteen: the 45th percentile. Before the typical crash, the tail-risk gauge was registering less tail risk than on an ordinary day.

Where SKEW stood at the peak before some of the worst drops since 1990. Most of the great crashes arrived with ordinary or low SKEW.

Market peak	Drop that followed	SKEW percentile then
Mar 2000 (dot-com top)	-11%	22nd
Oct 2007 (before the GFC)	-10%	47th
Jan 2009 (the final GFC leg)	-27%	29th
Jan 2018 (before Volmageddon)	-10%	41st
Jan 2020 (before COVID)	-16%	82nd
Mar 2022	-16%	90th
All 16 crashes (median)		45th

The list reads like an indictment. The dot-com top at the 22nd percentile. The 2007 top before the financial crisis at the 47th. The brutal final leg into early 2009, a 27% drop, at the 29th. The market's premier crash gauge was below its own median before nearly every one. The only crashes it partly flagged were a couple of recent ones, late 2021 and early 2025, when SKEW happened to be high for unrelated reasons. You cannot build a warning system on a gauge that is quiet before most of the events it is supposed to warn about.

And when it shouted, nothing happened

The other half of a predictor's job is not crying wolf, and here SKEW fails just as badly. After its loudest readings, the top 10%, a drop of more than 10% followed within two months only 10% of the time, actually lower than the 13% you would get on a random day.

The crash rate after SKEW's loudest warnings, versus a random day. The warning lowers the odds, it does not raise them.

Setup	Odds of a 10% drop in the next 60 days
After a top-decile SKEW reading ("the warning")	10%
Any day (baseline)	13%

Ninety percent of the loudest tail-risk warnings were followed by no crash at all. A fast SKEW spike, a sharp five-day jump, did no better, an 11% crash rate against the same 13% baseline. So SKEW misses in both directions at once: quiet before the crashes that came, loud before the crashes that never did.

Why a crash index cannot see crashes

The flaw is in what SKEW measures. It is not the probability of a crash; it is the price of crash insurance, the relative cost of out-of-the-money puts. And that price is driven by demand for hedges, which tends to be highest when markets are calm and grinding higher and investors are comfortable paying a little for protection. SKEW is often at its highest in exactly the complacent, melt-up conditions that precede no crash. Real crashes, by definition, are the ones the market did not price in, which is precisely why SKEW was ordinary before them. It is a sentiment and positioning gauge dressed up as a forecast.

It has also drifted. SKEW averaged 116 in the 1990s and 139 in the 2020s, ratcheting higher as the appetite for tail hedges has grown. A reading that once looked extreme is now routine, so a high SKEW today tells you even less than it used to. A signal that is always elevated is no signal at all.

How we tested it

• Data. The CBOE SKEW index daily from 1990 to today, 9,156 trading days, aligned with the S&P 500 over the same dates.
• Forward returns and drawdowns. The S&P's return and its worst drawdown over the next 20, 30, and 60 trading days, grouped by SKEW percentile bucket and by absolute SKEW level.
• Correlation. The straight-line relationship between SKEW and both the next month's return and the next sixty days' worst drawdown.
• False negatives. SKEW's percentile at every market peak that preceded a drop of more than 10% within 40 trading days, to see whether it was elevated before the actual crashes.
• False positives. The crash rate after SKEW's top-decile readings versus a random day, and the same for a sharp five-day SKEW spike.
• The drift. SKEW has trended structurally higher over the period, so we cut the data both by full-sample percentile and by absolute level to make sure the conclusion does not hinge on a moving baseline.