What Is Volatility Clustering? Why Calm Follows Calm

Put a volatility measure under almost any price chart and you do not get a flat line with the occasional random spike. You get something lumpier: long stretches of quiet, then a burst of turbulence that arrives all at once and takes weeks to drain away. Calm months look calm throughout. Chaotic months look chaotic throughout.

That lumpiness is not a quirk of one market or one era. It is one of the most durable descriptive patterns in financial data, and it has a name: volatility clustering. It is also the reason a whole family of chart tools — every band drawn from recent volatility, including ours — is able to work at all.

The pattern in one sentence

Volatility clustering is the tendency for large price changes to be followed by large price changes, and small changes to be followed by small changes — regardless of direction.

That last clause is the part people skip, and it is the part that matters most. Clustering is a statement about magnitude, not about sign. A market in the middle of a turbulent cluster is one where large moves have been arriving in succession. Whether the next one is big and up or big and down is a completely separate question, and clustering has nothing to say about it.

This is why volatility clustering is not a directional edge, and why no honest description of it ends with a trade. It narrows the question "how far might this move?" It leaves the question "which way?" exactly where it found it.

Why it is not the same as a trend

A trend is persistence in direction: up days following up days. Clustering is persistence in size: big days following big days, in either direction. The two are independent. A market can grind steadily upward in tiny increments — a strong trend with almost no volatility — or thrash violently sideways for a month and end where it started, which is a strong volatility cluster with no trend at all.

Keeping the two apart is what lets a volatility measure stay useful when direction is genuinely unknowable. It is also why a calibrated expected range describes width rather than destination.

Why every rolling volatility band depends on it

Here is the quiet assumption inside every band built from recent market behavior, from the simplest average true range to the most elaborate risk model.

To draw a range around tomorrow's price, a tool has to estimate how volatile tomorrow will be. It cannot observe tomorrow. So it looks backward — at the last 20 bars, or 60, or 200 — computes how much the market has been moving, and projects that forward. Rolling volatility is exactly this: a running measurement of recent movement, updated bar by bar.

That projection is only defensible if recent volatility carries information about near-term volatility. If turbulence arrived and vanished at random — if a wild week told you nothing whatsoever about the following Monday — then a backward-looking window would be measuring noise, and every band built on one would be decoration.

Volatility clustering is the reason the projection is not decoration. "Tomorrow will be roughly as volatile as the recent past" has historically been most of the answer, one bar out. That is a low bar for sophistication, and deliberately so: a published head-to-head found that rolling volatility, GARCH and exponentially weighted models were statistically indistinguishable on the one-step-ahead calibration task. The more elaborate models are built to describe how clustering decays over longer horizons. One bar out, there is little room for that extra machinery to matter.

The window length is where clustering meets a practical trade-off. Our own model uses a 60-bar volatility lookback — roughly a calendar quarter on daily bars. A shorter window reacts to a new cluster faster but is noisier; a longer window is smoother but lags through a regime change. That trade-off exists only because clusters have duration. If volatility changed at random every bar, no window length would help.

The evidence that the pattern is stable

Clustering being real is one claim. Clustering being stable enough to build on is a stronger and more testable one — and it is the claim a calibrated band actually stakes itself on.

The test is straightforward: build a range from a rolling volatility window, then count how often the next close landed inside it. If clustering is durable, that hit rate should hold up across wildly different market eras. In our working paper, the S&P 500 was scored this way on daily closes from 1928 to 2024 — 24,248 trading days — and the next close landed inside the inner band 71.2% of the time. Broken out decade by decade, containment stayed within a 68.7%–73.7% band across eleven calendar decades: the Great Depression, the war years, stagflation, the 1987 crash, 2008, COVID, the recent AI cycle. The full table is on the Proof page. These are historical measurements; past behavior is not a guarantee of future results.

Almost every other property of those markets changed beyond recognition — participants, technology, regulation, tick sizes, trading hours. The persistence of volatility was one of the few things that did not move much. That stability, rather than the exact percentage, is what makes the pattern worth naming.

Where the assumption breaks

Clustering is an average property, and averages have exceptions. The one that matters is the fast crisis.

When volatility jumps in a single session, a backward-looking window is still describing the calm that preceded it. The cluster has started; the estimate has not caught up. For a stretch of days the range runs too narrow, and closes land outside it more often than the long-run rate would suggest. We measure that shortfall and publish it rather than round it off — the numbers and the mechanism are set out in when volatility bands fail.

This is worth stating plainly, because it is the honest limit of the whole idea: clustering makes a rolling estimate reasonable most of the time, and the exceptions land disproportionately in the moments people care about most.

What clustering does not give you

It does not give direction. It is a magnitude pattern, and no amount of it resolves which way a market goes next.

It does not give timing. Knowing that clusters persist does not tell you when the current one ends. "Volatility is elevated and tends to stay elevated for a while" is a description of a tendency, not a countdown.

It does not convert into a forward probability for a particular chart. A containment rate measured across decades of history describes what happened in that sample. It is not a promise about the next bar on the instrument in front of you.

And it does not, by itself, make anything profitable. Our working paper validates the calibration of an expected range — how often price historically stayed inside it — not the profitability of any particular way of using it. Whether any such use delivers value after real-world costs is an open question, and trading carries risk, including the possible loss of capital.

The takeaway

Volatility clustering is the tendency for turbulence to arrive in runs rather than at random. It is a statement about how big moves are, not which way they go. It is the reason a rolling volatility window is a defensible way to estimate near-term range — and, when a cluster begins faster than the window can register, it is also the reason such estimates fail in exactly the way they do.

It is a useful thing to be able to name, because it turns a vague chart impression into something measurable. Whether any given tool has actually measured it is a separate question, and one worth asking of every band on your chart.

Frequently asked questions

Is volatility clustering the same thing as mean reversion? They describe different timescales and are usually complementary rather than opposed. Clustering describes short-run persistence — turbulent days grouping with turbulent days. Mean reversion in volatility describes the longer-run tendency for extreme calm or extreme turbulence to drift back toward a typical level eventually. Models like GARCH are built to represent both at once, which is one reason they carry more machinery than a plain rolling window; the published comparison found that machinery did not measurably help on the one-bar-ahead calibration task.

What causes volatility to cluster? There is no single settled answer, and we do not publish a causal claim of our own — our research measures the pattern rather than explaining its origin. Commonly discussed explanations involve the arrival of information in bursts and the way market participants respond to uncertainty, but these remain competing explanations rather than a demonstrated mechanism. The practical point does not depend on resolving it: the pattern is measurable whether or not its cause is agreed.

Why do volatility bands widen after a big move? Because the band's width is computed from recent volatility, and a large move raises that measurement as it enters the lookback window. The widening is the estimate updating, not a forecast that more turbulence is coming. It also means the widening arrives after the move that caused it, which is the source of the crisis-onset lag described here. What each line on a band is actually reporting is covered in reading the expected range.

Does volatility cluster in crypto and forex the same way it does in stocks? The containment test that leans on clustering held up across equities, FX, commodities, rates and crypto in our sample, with crypto sitting slightly higher at the inner band than the equity rows. The per-instrument figures are tabulated on the Proof page. Those are historical measurements on the instruments tested, and past behavior is not a guarantee of future results; markets outside that universe carry the usual caveats.

If you want to see a volatility estimate update in real time — widening as a cluster begins, tightening as one fades — the most direct way is to watch it recalculate on the markets you already follow. BTM draws a calibrated expected range on any TradingView chart, and you can start a free 30-day trial to put it next to whatever you use now.