Volatility clustering
Calm days tend to follow calm days, and wild days follow wild days, so recent swings say a lot about the next few.
5 min readLesson 18 of 57Interactive
In short
A GARCH model turns that into a forecast: tomorrow's variance is a long-run level, plus a share of today's surprise, plus a share of today's variance. After a shock, the forecast is high and then fades back towards normal. It forecasts how big moves are likely to be, never which way they'll go.
Why it matters
Markets don't swing by the same amount every day. They have moods: weeks of small, quiet moves, then a shock, then weeks where 2% and 3% days are ordinary. A stormy day makes the next few more likely to be stormy too, and calm tends to follow calm. This is volatility clustering, one of the most reliable patterns in market prices.
It matters because it makes risk partly forecastable, even when direction isn't. Yesterday's chaos says little about whether tomorrow goes up or down, but a lot about how far it's likely to move. That helps when sizing a position, setting a stop, judging a Value at Risk figure, or deciding whether a calm stock's sudden jumpiness is unusual.
How it works
The standard model is GARCH(1,1). It forecasts tomorrow's variance, which is volatility squared, from three parts: a small share of the long-run level, a share called alpha of today's squared surprise (how far today's move was from normal), and a share called beta of today's variance. A big surprise lifts the forecast straight away; beta makes the higher level linger.
Alpha plus beta is the persistence: the share of a shock's extra variance still there the next day. Stock indexes usually land around 0.97 to 0.99, so shocks fade slowly. The half-life is how many sessions it takes for half of a shock to fade: log(0.5) ÷ log(alpha + beta). At a persistence of 1 or more, shocks would never fade and there'd be no long-run level.
The forecast is about size, never direction. From it come price ranges: if tomorrow's volatility is 1.5%, the next move lands within ±1.5% on about 2 days in 3, and within ±1.645 × 1.5% on about 9 days in 10. Those ranges assume moves follow a bell curve, and real markets don't: crashes happen far more often than a bell curve allows, so the outer edges understate them.
The volatility cone is a second view. Measure volatility over the last week, month, three months, six months and year, each as a yearly rate, and compare each with its own history: the band from the 10th to the 90th percentile of past readings. Above the band, today is unusually stormy for that window; below it, unusually calm. Short windows swing more than long ones, so the band is wider at the short end, like a cone.
MarketLab's simulated market is itself built with GARCH(1,1): the market factor uses an alpha of 0.09 and a beta of 0.89. So the model fits these prices unusually well, and fitting it to the simulated S&P 500 should find numbers close to those. That's a lesson in itself: a model finds its pattern easily when the data was made that way. Real markets are messier, with jumps, crashes and moods that change shape.
variance tomorrow = (1 − alpha − beta) × long-run variance + alpha × today's surprise² + beta × today's variance; half-life = log(0.5) ÷ log(alpha + beta)
Tomorrow's risk is mostly today's risk, nudged by today's surprise and pulled slowly back towards normal.
A worked example
A stock's long-run daily volatility is 1% (about 15.9% a year), with an alpha of 0.09 and a beta of 0.89. Today's volatility was the usual 1%, but the stock fell 4%, closing at $100.
- Work in squared percentages. The long-run variance is 1² = 1, and its share is (1 − 0.09 − 0.89) × 1 = 0.02.
- Today's surprise adds 0.09 × 4² = 0.09 × 16 = 1.44. Today's variance adds 0.89 × 1 = 0.89.
- Tomorrow's variance is 0.02 + 1.44 + 0.89 = 2.35, so tomorrow's volatility is √2.35 ≈ 1.53%, or 1.53% × 15.87 ≈ 24.3% a year.
- Ranges for tomorrow: about 2 in 3 between $98.47 and $101.53 (±1.53%), and 9 in 10 within ±1.645 × 1.53% ≈ ±2.52%, from $97.48 to $102.52.
- Fading: the persistence is 0.09 + 0.89 = 0.98, so the half-life is log(0.5) ÷ log(0.98) ≈ 34 sessions, about seven weeks.
Result: One 4% fall lifted tomorrow's expected swing by about half, from 1% to 1.53%, and it takes about 34 sessions for half of that extra to fade. The model says nothing about whether the next move is up or down.
Try it
Calm and wild spells
Two years of daily moves with 19% volatility overall. The calmest month ran at 9% a year and the wildest at 37%.
Volatility overall
19%
Two years together
Calmest month
9%
As a yearly rate
Wildest month
37%
As a yearly rate
Biggest day
−4.8%
A shock fades by half in
34 days
log 0.5 ÷ log(alpha + beta)
Generated prices for illustration, not market data.
How to read it
- Volatility now well above its long-run level
- A stormy spell. Ranges are wide, and the model expects them to narrow only slowly.
- Volatility now below its long-run level
- A calm spell. Calm usually follows calm, but not forever, and it's when risk numbers built from recent days look smallest.
- Half-life of a few sessions
- Shocks fade fast. Recent days matter, but not for long.
- Half-life of a month or more
- Spells last. Typical of stock indexes, and of MarketLab's simulated market.
- Today above the cone's band
- Volatility over that window is higher than in 90% of past readings.
In the real world
- Benoit Mandelbrot noticed in 1963 that large changes tend to be followed by large changes, of either sign, and small by small. Robert Engle turned that into the ARCH model in 1982, Tim Bollerslev extended it to GARCH in 1986, and Engle shared the 2003 Nobel prize in economics for the work.
- Banks and funds use GARCH-style forecasts for risk limits and Value at Risk, and options traders compare them with implied volatility. Some funds target a set volatility, holding less when the forecast rises.
- In 2008 and again in 2020, the S&P 500's daily swings stayed huge for weeks before settling, a cluster you can see in any chart of its daily moves. October 19, 1987, when it fell about 20% in one day, would be close to impossible under a bell curve.
Common mistakes
- Reading a high volatility forecast as a prediction of a fall. It means big moves, in either direction.
- Treating the 9-in-10 range as the worst case. On about 1 day in 10 the move lands outside it, and crashes can land far outside.
- Using one long-run volatility number just after a shock or a long calm, the two times it's most wrong.
- Trusting a fit because it looks perfect on simulated prices. MarketLab's were made with this model; real ones weren't.
See it in MarketLab
How much it might move →
On a stock page's Risk & technicals tab, this card shows likely price ranges for the next day, week and month (about 2 in 3, and 9 in 10), volatility now against its long-run level, how fast a shock fades by half, and the volatility cone.
Did the forecast work? →
The same card checks the last 250 sessions, with a model fitted only on the days before each one: the share of days inside the 90% range, next to a plain last-month's-volatility forecast. Close to 90% means the ranges were the right size.
Test yourself
Optional: 4 quick questions on the words, using the idea, and how it plays out in real markets.
Loading your answers…