Plot the squared daily returns of any major index or currency pair and the pattern jumps out at you. Calm stretches and turbulent stretches take turns. Big moves, up or down, tend to be followed by more big moves. Small moves get followed by small moves. This is one of the most reliable empirical facts in markets, and it is not subtle. You do not need a model to see it.
Benoit Mandelbrot spotted it in cotton prices back in the 1960s. Robert Engle turned it into math with the ARCH model in 1982, and Tim Bollerslev extended it into GARCH in 1986. The point of all these models is the same. Today's volatility is partly predictable from yesterday's. Volatile yesterday means more likely volatile today. Quiet yesterday means more likely quiet today. The link is not perfect, but it is strong enough to actually trade around.
Why clustering happens
Clustering is really about how news arrives and how long markets take to digest it. A surprising Fed decision does not create one bad day and then stop. It kicks off a stretch of elevated uncertainty while traders reassess positions, analysts rebuild their models, and fresh information about the policy shift keeps trickling in. The uncertainty persists because the disagreement persists.
Then feedback makes it worse. When vol rises, risk models across the system start flashing. Hedge funds and bank desks cut positions to stay inside their limits. That forced selling, or buying back of shorts, creates more price movement, which shows up as more volatility, which trips more risk reduction. That self-reinforcing loop is why vol tends to spike fast and hang around before slowly bleeding off.
Microstructure adds to it too. When things get choppy, market makers widen their spreads to get paid for the risk of holding inventory. Wider spreads mean even normal-sized orders push prices around more, which reads as higher volatility in the data. Liquidity pulling back during stress is a structural amplifier, so clustering ends up more pronounced than the raw information flow alone would justify.
Putting numbers on it
The simplest way to measure the effect is the autocorrelation of squared or absolute returns. For daily S&P 500 returns, the lag-1 autocorrelation of squared returns usually sits around 0.2 to 0.4, so roughly 20 to 40 percent of today's volatility is predictable from yesterday's. And it decays slowly. It stays statistically significant for weeks, sometimes months, which is exactly why vol regimes can drag on so long.
GARCH models pin this down formally. The standard GARCH(1,1) estimates tomorrow's conditional variance as a weighted blend of the long-run average variance, today's squared return, and today's conditional variance. Those weights tell you how fast vol reverts to its mean and how sharply it reacts to new shocks. For equity indices the mean reversion usually comes out slow, which just confirms that regimes are sticky.
What it means for risk management
If you size positions off a fixed volatility estimate, you are wrong in a predictable way. In quiet periods your estimate runs too high, so you hold less than you could. In wild periods it runs too low, so you hold more than you should. Both cost you, but the second one is the dangerous one, because it leaves you overexposed right when the market is most likely to hand you a large adverse move.
Adaptive sizing fixes this head-on. Instead of a fixed number, you feed a recent volatility measure into the sizing, something like a 20-day exponentially weighted moving average of absolute returns. Low vol means bigger position, high vol means smaller position. Your dollar risk stays roughly constant across environments instead of ballooning whenever the tape gets rough. This is the same logic we bake into position sizing on Blockcircle, and it is one of the highest-leverage habits a discretionary trader can pick up.
Stops need the same treatment. A stop 2 percent below entry might be fine in a calm regime and then get run over by ordinary noise once vol picks up. Set stops at a multiple of recent volatility instead, say 2x the 20-day average true range, and they widen and tighten with the regime so you are not getting knocked out by ordinary chop.
Reading the regime
Knowing whether you are in a high-vol or low-vol regime is worth more than just sizing. In high-vol regimes, cross-asset correlations tend to climb, so diversification quietly stops helping. Trend signals get flakier and mean-reversion strategies tend to do better. In calm regimes it usually flips the other way.
You can get a decent read cheaply. Compare current realized vol to its longer-term average. If 20-day realized vol is more than one standard deviation above its 1-year average, call it a high-vol regime. The VIX works as a live proxy for the equity vol regime too. Above 25 generally lines up with high-vol conditions, below 15 with calm ones.
Fancier setups use hidden Markov or regime-switching models to actually estimate the probability of being in each state, and they can catch turning points faster than a simple threshold. They are also more work to build and calibrate. For most real trading decisions, comparing current vol to its recent average gets you most of the value without the overhead, so start there and only reach for the heavy machinery if you have a concrete reason to.