Volatility Clustering

Pull up a chart of daily returns on almost any traded asset — a stock, an index, a currency, a cryptocurrency — and a texture leaps out that no textbook of independent draws would ever produce. For weeks the returns are placid: small green and red bars, barely a flicker. Then something breaks — an earnings miss, a rate decision, a panic — and suddenly the bars explode in both directions and stay large for days or weeks before calming down again. Turbulence begets turbulence; quiet begets quiet. This is volatility clustering, and it is the single most important empirical fact about financial markets.

Benoit Mandelbrot named it in 1963, watching cotton prices: "large changes tend to be followed by large changes — of either sign — and small changes tend to be followed by small changes." That one sentence quietly demolishes the comfortable assumption that returns are white noise with a fixed variance. The level of a return may be unpredictable, but its size is not.

The two stylized facts that break white noise

Let r_t be the return (say the log-return \log P_t - \log P_{t-1}) on day t. Decades of data across every liquid market agree on a short list of stylized facts. Two of them are the heart of this module, and they pull in opposite directions:

Returns are almost uncorrelated. The autocorrelation of r_t itself is negligible at every lag: \operatorname{Corr}(r_t, r_{t-k}) \approx 0 for k \ge 1. You cannot predict the sign or direction of tomorrow's return from the past — markets are (nearly) unforecastable in the mean.
Yet returns are clearly not independent. The autocorrelation of the squared (or absolute) returns is strongly positive and decays only slowly: \operatorname{Corr}(r_t^2, r_{t-k}^2) > 0 out to many lags. Today's magnitude is very predictable from yesterday's magnitude.

Read those two lines together and the paradox is stark: the sequence is serially uncorrelated but very far from independent. All the dependence has fled from the returns and hidden in their squares — in the variance. A model that gets the mean right but treats the variance as a constant has thrown away the most exploitable structure in the whole series.

Seeing the clusters

Below is a stylized daily-returns series. Notice how it hovers around zero throughout — there is no trend, no drift, the mean really is flat. What changes is the amplitude. The two shaded windows are turbulent: the swings are three to five times wider than in the calm stretches on either side, and — crucially — a big day sits next to other big days rather than being sprinkled at random. That clumping is volatility clustering made visible.

Shuffle this series in time and the mean, the histogram, and the (near-zero) autocorrelation of r_t would all be unchanged — but the clusters would vanish, smeared into a uniform fuzz. The clustering lives entirely in the ordering of the magnitudes, which is exactly the dependence a variance model must capture.

The fingerprint: two autocorrelation functions

The cleanest diagnostic is to plot the sample autocorrelation function twice — once for r_t and once for r_t^2 — on the same axes. The contrast is the signature of every financial return series ever measured:

The returns' ACF (near zero at every lag) says "no linear forecastability." The squared returns' ACF, decaying slowly like the ACF of a persistent stationary process, says "the variance has a long memory." Whenever you see this pattern, reach for a conditional variance model: an ARCH or GARCH.

Fat tails come along for free

A third stylized fact rides on the back of clustering: fat tails, or leptokurtosis. If you histogram the returns, the tails are heavier and the peak is taller than a normal (Gaussian) bell of the same variance — extreme days (a −20% crash) happen far more often than a normal law would ever allow. This is not a separate mystery. A process that is Gaussian on any single day but whose variance itself changes over time is a mixture of narrow and wide bells, and any such mixture of normals has excess kurtosis — fatter tails than a single normal. Clustering and fat tails are two views of the same fact: the conditional variance is not constant.

A quick worked check. Suppose returns are Gaussian with standard deviation 0.5\% on a calm day and 2.5\% on a turbulent day, split half and half. The overall variance is \tfrac12(0.5)^2 + \tfrac12(2.5)^2 = 3.25\ (\%^2), so an overall standard deviation of 1.80\%. But a genuine single normal with that same 1.80\% deviation would essentially never produce an 8\% move — whereas on the turbulent regime an 8\% day is only a bit beyond three standard deviations, entirely plausible. The mixture manufactures the heavy tail without anything exotic per day.

This is the misconception the whole module exists to correct. For jointly normal variables, zero correlation does imply independence — and that special case has trained generations of students to treat the two words as synonyms. Financial returns are the loud counterexample. They are (almost) perfectly uncorrelated: no linear relationship links r_t to its past, so linear forecasts of the return are useless. Yet they are emphatically dependent — the nonlinear function r_t^2 is strongly autocorrelated, so the past tells you a great deal about tomorrow's size. Correlation only ever measures the linear part of dependence; a variable can carry huge information about the spread of another while carrying none about its mean. Mistaking "uncorrelated" for "independent" is precisely what makes a constant-variance model look adequate when it is throwing away the most valuable structure in the data.

There is real economics under the statistics. Information arrives in bursts, not on a smooth schedule — a central-bank surprise, a war, a bankruptcy — and it takes markets several sessions to digest a shock, argue about its meaning, and unwind leveraged positions, so a jolt echoes for days. Traders also react to each other: a big move raises everyone's estimate of risk, margin calls force more selling, and rising uncertainty feeds on itself. The result is that the effective "temperature" of the market — its volatility — is a slow-moving, persistent state, even though the direction of each day's move stays a coin flip. Modelling that persistent hidden temperature is exactly what ARCH and GARCH do.