4CSDS3021: Time Series Analysis
Module 3: Modeling Volatility Clustering
Lecture Notes | Hours allotted: 9
Topics covered
● Volatility: meaning and measurement
● Volatility clustering
● Econometric models of volatility
● The ARCH model
● The GARCH model and its extensions
● Testing for ARCH/GARCH effects
1. Volatility: Meaning and Measurement
Volatility refers to the dispersion, or variability, of a series around its mean — most commonly discussed in the context
of asset returns, where it is a proxy for risk. Unlike the AR/MA models of Module 1, which model the conditional
MEAN, volatility models focus on the conditional VARIANCE.
1.1 Common measures
● Historical (unconditional) standard deviation of returns over a chosen window
● Conditional variance from a fitted ARCH/GARCH-type model — varies over time
● Implied volatility — backed out from option prices (e.g. the VIX)
● Realized volatility — computed from high-frequency intraday data
2. Volatility Clustering
Volatility clustering is the empirical regularity, first noted by Mandelbrot, that 'large changes tend to be followed by
large changes, of either sign, and small changes tend to be followed by small changes.' Periods of high volatility and
periods of calm each tend to persist.
Implication
Although returns themselves are typically close to uncorrelated (hard to predict in the mean), the SQUARED or
absolute returns are strongly autocorrelated. This means the conditional variance is predictable even when the
conditional mean is not — exactly what ARCH/GARCH models exploit.
2.1 Other stylised facts of financial returns
● Leptokurtosis (fat tails): extreme returns occur more often than a normal distribution would predict
● Leverage effect: volatility tends to rise more after negative shocks than after positive shocks of the same size
● Volatility mean reversion: shocks to volatility eventually decay back toward a long-run average level
3. Econometric Models of Volatility
Classical regression assumes homoskedasticity — constant error variance. Volatility clustering is evidence of conditional
heteroskedasticity: the variance of the error, conditional on past information, changes over time even if the
UNconditional variance is constant.
Xt = μt + εt , εt = zt σt , zt ~ i.i.d.(0,1)
● μt is the conditional mean (an ARMA specification, as in Modules 1-2)
● σt² is the conditional variance — modelled as a function of past information
● zt is a standardised innovation, typically assumed Normal or Student-t (to allow fat tails)
4. The ARCH Model
Engle's (1982) Autoregressive Conditional Heteroskedasticity (ARCH) model lets the conditional variance depend on
past squared shocks:
σt² = ω + α1 εt−1² + α2 εt−2² + ... + αq εt−q²
● ω > 0, and αi ≥ 0 for all i, to guarantee σt² > 0 at every t
● A large shock εt−1² in the recent past raises today's conditional variance — this is how ARCH generates
volatility clustering
4.1 Stationarity condition
● The ARCH(q) process is covariance-stationary (in the sense of having a finite, constant unconditional variance)
iff Σ αi < 1
● Unconditional variance: σ² = ω / (1 − Σ αi)
4.2 A practical limitation
Capturing realistic, gradually-decaying volatility persistence often requires a very high ARCH order q, which is
parameter-heavy and can produce negative variance estimates unless heavily constrained. This motivated the GARCH
extension.
5. The GARCH Model and Extensions
5.1 GARCH(p, q) — Bollerslev (1986)
Generalized ARCH adds lagged conditional variances to the ARCH specification, giving an ARMA-like structure to the
variance equation:
σt² = ω + Σ (i=1..q) αi εt−i² + Σ (j=1..p) βj σt−j²
● GARCH(1,1) — σt² = ω + α εt−1² + β σt−1² — is by far the most widely used specification in practice, and
often outperforms high-order ARCH models with far fewer parameters
● Stationarity: α + β < 1; unconditional variance = ω / (1 − α − β)
● α + β close to 1 indicates highly persistent volatility shocks (near-integrated variance, sometimes modelled with
IGARCH)
5.2 Key extensions
Model Captures
EGARCH Exponential GARCH — models log(σt²), so no non-
negativity constraints on parameters are needed; captures
the leverage effect (asymmetric response to +/- shocks)
GJR-GARCH / TGARCH Adds a term active only for negative shocks, directly
capturing the leverage effect with a threshold indicator
IGARCH Integrated GARCH — imposes α + β = 1: shocks to
volatility are permanent (infinitely persistent)
GARCH-M GARCH-in-Mean — feeds the conditional variance (or
SD) directly into the mean equation, letting expected
return depend on risk
6. Testing for ARCH/GARCH Effects
Before fitting an ARCH/GARCH model, it is standard practice to test whether conditional heteroskedasticity is actually
present in the residuals of a fitted mean equation.
6.1 Engle's ARCH-LM test
Fit the mean equation (e.g. ARMA) and obtain residuals ε̂t. Then regress the squared residuals on their own lags:
ε̂t² = α0 + α1 ε̂t−1² + ... + αq ε̂t−q² + vt
● H0: α1 = α2 = ... = αq = 0 (no ARCH effects — homoskedastic)
● Test statistic: T·R² from this auxiliary regression ~ χ²(q) under H0 (T = sample size)
● A significant test statistic (reject H0) is evidence that an ARCH/GARCH model is warranted
6.2 Ljung-Box test on squared residuals
● Alternative/complementary check: apply the Ljung-Box portmanteau test (Module 1) to the squared
standardised residuals; significant autocorrelation again signals ARCH effects
Model-building workflow
1) Fit an appropriate ARMA mean equation and obtain residuals. 2) Test the residuals for ARCH effects (ARCH-LM
/ Ljung-Box on squared residuals). 3) If present, jointly estimate the mean and a GARCH-type variance equation by
maximum likelihood. 4) Check standardised residuals for remaining ARCH effects and evaluate fit.
Summary
Module 3 in one paragraph
Volatility clustering — the tendency of large and small changes to each cluster together — shows up as strong
autocorrelation in squared or absolute returns even when returns themselves are nearly uncorrelated. ARCH models
let today's conditional variance depend on past squared shocks; GARCH generalises this by also including lagged
conditional variances, with GARCH(1,1) as the workhorse specification. Extensions such as EGARCH and GJR-
GARCH capture the leverage effect (asymmetric volatility response to positive vs. negative shocks), while IGARCH
models permanent volatility persistence. Before fitting any of these, the ARCH-LM test (or Ljung-Box on squared
residuals) is used to confirm that conditional heteroskedasticity is actually present.