What Is Volatility Clustering? 76 Years of S&P 500 Data

Michael Whitman · Trading since 2017, full-time since 2020 · YouTube creator, 137K subscribers · Published 2026-09-14 · Updated 2026-10-27 · Reviewed 2026-10-27

Volatility clustering is the tendency for large price moves to arrive next to other large moves, and small ones next to small ones, with no lean in direction. On the S&P 500 from 1950 to October 2026, a day in the loudest tenth was followed by another 24.12% of the time, against 8.43% after any other day.

If you had to keep one statistical fact about markets and throw away the rest, this would be a strong candidate. It is also the fact most likely to be misread as a trading signal, which it is not.

How it works

Large moves are followed by more large moves. Not always, but far more often than chance would produce. A violent day makes the next day more likely to be violent.

It shows up in every decade of the index’s record. The link between one day’s size and the next was positive in all eight decades from the 1950s to the 2020s, which is why volatility models are built around it and direction models mostly are not.

It is stronger on the loud side than the quiet side. A calm day is followed by another calm day only a little more often than chance, so a long quiet stretch is less of a regime than it feels like from inside it.

And it has nothing in it about direction. A cluster of large bars can be a crash or a melt-up. The property predicts magnitude and is silent on sign.

A worked example

Take a fixed 1% stop on the S&P 500 and two real stretches. In 2017 the index had 251 sessions and only 8 of them closed 1% or more away from the previous close; the median day moved 0.185%. In March and April 2020 it had 43 sessions and 37 of them moved 1% or more; the median day moved 2.93%.

In 2017 the stop sat more than five median days away. A 1% distance divided by a 0.185% median day is about 5.4, so ordinary movement rarely reached it, and the largest close-to-close change of the year was 1.82%, on 17 May.

In spring 2020 the same stop was about a third of a median day. Ordinary movement covered it on 37 of 43 sessions, so a position using it was being decided by noise rather than by the idea behind the trade.

Same rule, same index, opposite behavior, and the switch between the two regimes is not random. It is clustered, which means a bad stretch is a stretch rather than a bad day.

What it is good for

Sizing, and only sizing. Because the property predicts magnitude, it supports one useful action: adjusting how much room you leave and therefore how large a position you hold.

That is what ATR operationalizes. A stop set at a multiple of recent average range widens in a loud cluster and tightens in a quiet one, because the input is itself measuring the cluster.

It is not a reason to trade. An expansion in volatility is not a directional signal, and the most common misuse of this property is treating “something is happening” as “I know what.”

Why it exists

The honest answer is that nobody fully knows, and several plausible mechanisms all contribute.

Information arrives in clusters. News does not distribute evenly through time; announcements, earnings and policy decisions bunch together, and markets respond while they are being digested rather than instantly.

Positions unwind in sequence. A large move forces some participants to reduce, that reduction moves price further, which forces the next round. The mechanism is mechanical rather than informational and it takes days rather than seconds.

And liquidity withdraws when it is most needed. Market makers widen or step back during turbulence, so the same order size moves price further, which produces more turbulence, which keeps them away.

All three are self-reinforcing, and none of them is directional. Each explains why large moves beget large moves without offering any view on which way, and that is the structural reason the property is so dependable about magnitude and so useless about sign.

Size has memory, sign does not

The cleanest way to see clustering is to compare two correlations. Take every daily change and ask how strongly it is linked to the change one, two or 250 sessions earlier, once for the size of the move and once for its signed value.

For size the link is positive at every lag tested. It starts at 0.2688 one session apart, is still 0.1953 twenty sessions apart, 0.0937 a hundred apart and 0.0492 a full year apart. All 250 values sit above the band that pure chance would produce on a sample this long.

For the signed move the link is close to zero almost everywhere. One session apart it is -0.0018, and 39 of the 250 lags fall outside the chance band, none by more than 0.0367 (at 16 sessions). Yesterday’s direction says almost nothing about today’s; yesterday’s size says a good deal about today’s.

The original data

Sample: S&P 500 daily closes from Yahoo Finance (^GSPC), 3 Jan 1950 to 23 Oct 2026, giving 19,324 daily changes measured as log changes from one close to the next. Every gap of more than four calendar days in the file is a known market closure, and the 1950s carry 67 unchanged closes.

Loud and quiet days. A loud day is one in the top tenth by size, a change of 1.4695% or more (1,932 days). The next day was loud too 24.12% of the time, against 8.43% after the other days, and it closed up 52.64% of the time, against 53.12% for all days. A quiet day is one in the bottom tenth, 0.0782% or less (1,932 days); 11.90% were followed by another. Those cut-offs use all 76 years, later days included. Set instead from only the 1,000 sessions before each day (from January 1954), a loud day was followed by another 24.39% of the time, against 9.23% after the other days.

By decade, one session apart, the size link was 0.1700 in the 1950s, 0.3267 in the 1960s, 0.2126 in the 1970s, 0.1881 in the 1980s, 0.1530 in the 1990s, 0.2427 in the 2000s, 0.2134 in the 2010s and 0.3695 from 2020 to 2026.

Second source. FRED’s S&P 500 series covers 27 Oct 2016 to 23 Oct 2026; its 96 blank dates are market holidays, which Yahoo’s file also skips, so both give 2,510 changes. On that window the size link one session apart is 0.3702 on both, and 20 sessions apart 0.1946 on FRED against 0.1945 on Yahoo. The lags are in the S&P 500 autocorrelation file, and the loud and quiet counts in the follow-through file.

Real drawdowns are the loud stretches. The index slid 5% or more below its record 68 separate times between 1950 and September 2026, and fell 56.8% between 9 October 2007 and 9 March 2009: not one bad day but a long run of them, in which a stop sized on the calm year before was the wrong size day after day.

A fixed cost is a different share in each regime. A hypothetical round-trip cost of 0.02% of price is 10.81% of 2017’s median daily move and 0.68% of the median move in March and April 2020.

When it fails

The characteristic failure is mistaking persistence for prediction. Clustering says the next bars are likely to be large, and that feels like foreknowledge. It is a statement about the size of the distribution, not its center: after the loudest tenth of days, the next close was up about as often as on any day. A position taken because “volatility is picking up” has been taken on information that contains no direction at all.

A second failure is assuming a quiet regime will continue. Quiet days repeat only a little more often than chance, and the transition out is usually abrupt rather than gradual.

A third is sizing up during a calm stretch because recent losses have been small, which raises exposure precisely as the quiet regime ages.

A fourth is using a fixed percentage stop across regimes, which risks a different multiple of ordinary movement in each one, as 2017 and spring 2020 show.

And a fifth is expecting a loud stretch to end on schedule. The size link fades slowly, from 0.2688 one session apart to 0.0937 a hundred sessions apart, so no fixed number of days marks the end of one.

Questions people ask

What does volatility clustering mean?

That large moves bunch together and small moves bunch together. On the S&P 500 since 1950, a top-tenth day was followed by another 24.12% of the time, against 8.43% after other days.

Does volatility clustering predict direction?

No. After a top-tenth day the index closed up the next day 52.64% of the time, against 53.12% for all days, and the signed move one session apart had a correlation of -0.0018.

How long does volatility clustering last?

Longer than most people expect. The link between move sizes was still 0.0937 a hundred sessions apart and 0.0492 a year apart, and it was above the chance band at all 250 lags measured.

How do traders use volatility clustering?

To size positions and stops, not to pick direction: a stop that sat more than five median days away in 2017 was about a third of a median day in spring 2020.

Related

Volatility covers the measure this property describes the behavior of. ATR covers the practical tool that adapts to it. And risk management covers sizing against a number that changes by regime.

This page is educational, not financial advice. Test every idea on your own charts before risking money.