Calm and storm, why volatility clusters
Returns are nearly unpredictable, but their size is not. GARCH(1,1) explains volatility clustering and fat tails with two parameters, and turns a single number for risk into a forecast.
The GBM and Kelly notes assume one fixed volatility . Look at any price chart and the assumption fails at a glance: quiet weeks follow quiet weeks, and a violent day is usually followed by more violent days. Mandelbrot noticed it in cotton prices in 1963: large changes tend to be followed by large changes, of either sign [1]. It is called volatility clustering, and it is one of a handful of “stylized facts” that hold across almost every liquid market [2]:
- daily returns are close to uncorrelated: yesterday’s sign says little about today’s;
- the size of returns ( or ) is strongly and persistently correlated;
- the distribution of returns has fatter tails than a normal one.
The first fact is why prices look like random walks. The second is why risk is forecastable even when direction is not.
GARCH(1,1)
Engle’s ARCH model [3] and Bollerslev’s generalization, GARCH [4], make tomorrow’s variance a function of today’s news. Write the daily return as with an independent standard normal draw, and let
A large move yesterday ( term) raises today’s variance, and the variance itself is sticky ( term). With the process settles around a long-run variance, and every shock fades geometrically:
The sum is the persistence. For daily data on most assets it lands between 0.9 and 0.99.
Fat tails for free
Every is normal, yet the returns are not. Mixing normals with different variances fattens the tails, and GARCH mixes them in a specific way. For normal shocks the kurtosis is
and the autocorrelation of squared returns starts at
and then decays as .
Take , and a long-run volatility of 60% a year (a 365-day year, since crypto trades every day). Then:
| Quantity | Theory | Simulation (200,000 days) |
|---|---|---|
| Shock half-life | 13.5 days | |
| Kurtosis | 3.77 | 3.75 |
| Autocorrelation of , lag 1 | 0 | −0.0005 |
| Autocorrelation of , lags 1 / 5 / 20 | 0.179 / 0.146 / 0.068 | 0.180 / 0.144 / 0.067 |
Two parameters reproduce all three stylized facts. They do not reproduce the amount of tail seen in real data, which is why practical fits usually replace the normal with a Student-.
The same facts in Bitcoin
The “BTC daily” tab uses Binance’s BTCUSDT perpetual, daily closes from September 2021 to October 2026: 1,878 returns with an annualized volatility of 52%. All three facts are there:
- autocorrelation of returns at lags 1, 5 and 20 is −0.02, 0.02 and −0.04, inside the ±0.05 band that pure noise would produce with this many points;
- autocorrelation of squared returns is 0.18 at lag 1, 0.07 at lag 5 and zero by lag 20;
- kurtosis is 7.3, well above the normal 3.
The maximum-likelihood fit with normal shocks gives , : persistence 0.93, a shock half-life of about 10 days, and a long-run volatility of 54%. It reproduces the lag-1 correlation of squared returns almost exactly (model 0.18) but expects more memory further out (0.14 at lag 5 against 0.07 in the data), and its kurtosis is 3.7 against 7.3. After dividing each return by its fitted , the leftovers still have kurtosis 6.2: the clustering is explained, the tails are not.
Swapping the normal shock for a Student- fixes the tails with a striking estimate: about 3 degrees of freedom, barely enough for the variance to exist. The persistence then rises to 0.997, so in that fit shocks fade over months rather than days and the long-run level is poorly pinned down.
Fitting it
The model gives the likelihood directly: given the parameters, run the variance recursion over the data and add up the normal log-densities. Maximizing that is a three-parameter optimization.
import numpy as npfrom scipy.optimize import minimize
def simulate_garch(omega, alpha, beta, n, seed=11): rng = np.random.default_rng(seed) r, var = np.empty(n), np.empty(n) var[0] = omega / (1 - alpha - beta) for t in range(n): if t > 0: var[t] = omega + alpha * r[t - 1] ** 2 + beta * var[t - 1] r[t] = np.sqrt(var[t]) * rng.standard_normal() return r
def neg_log_likelihood(params, r): omega, alpha, beta = params if omega <= 0 or alpha < 0 or beta < 0 or alpha + beta >= 1: return np.inf var = np.empty_like(r) var[0] = r.var() for t in range(1, len(r)): var[t] = omega + alpha * r[t - 1] ** 2 + beta * var[t - 1] return 0.5 * np.sum(np.log(var) + r**2 / var)
long_run_daily_var = 0.60**2 / 365omega = long_run_daily_var * (1 - 0.10 - 0.85)r = simulate_garch(omega, 0.10, 0.85, n=2_000)fit = minimize(neg_log_likelihood, x0=[r.var() * 0.1, 0.05, 0.90], args=(r,), method="Nelder-Mead")print(fit.x)On 2,000 simulated days the fit returns , and a long-run volatility of
61%, close to the truth but not exact; individual parameters are noisier than their sum. For real work, the
arch package [5] does this with standard errors and other shock distributions.
Volatility as a forecast
The payoff is a forecast of variance at every horizon. Iterating the recursion,
so after a shock, near-term volatility is high and the forecast slides back to the long-run level with the half-life above. That shape, a term structure of volatility, is what risk limits, margin and option prices are really about.
It also closes a loop with the Kelly note. The growth-optimal leverage is ; if moves, so should position size. Scaling exposure by forecast volatility, “volatility targeting”, is Kelly with a time-varying and an honest admission that is unknown.
Takeaway
Direction is close to noise; magnitude is not. A model with two numbers, how hard news hits and how long it lingers, explains why calm and storm come in runs, why tails are fat even with normal shocks, and gives a risk forecast you can size positions with.
Data
Binance Data Collection, USDⓈ-M futures daily klines (BTCUSDT), data.binance.vision.
References
- B. Mandelbrot, “The variation of certain speculative prices,” Journal of Business 36(4), 394–419 (1963). doi:10.1086/294632
- R. Cont, “Empirical properties of asset returns: stylized facts and statistical issues,” Quantitative Finance 1(2), 223–236 (2001). doi:10.1080/713665670
- R. F. Engle, “Autoregressive conditional heteroscedasticity with estimates of the variance of United Kingdom inflation,” Econometrica 50(4), 987–1007 (1982). doi:10.2307/1912773
- T. Bollerslev, “Generalized autoregressive conditional heteroskedasticity,” Journal of Econometrics 31(3), 307–327 (1986). doi:10.1016/0304-4076(86)90063-1
- K. Sheppard et al., arch: ARCH models in Python. github.com/bashtage/arch, doi:10.5281/zenodo.593254