Bollerslev (1987) A Conditionally Heteroskedastic Time Series Model for Speculative Prices and Rates of Return

Summary

This short paper introduces the Student-$t$ GARCH (GARCH-$t$) model — a GARCH process whose standardized innovations follow a Student-$t$ distribution rather than a Gaussian. The motivation is that speculative returns are serially uncorrelated but exhibit volatility clustering and unconditional fat tails; a normal-error GARCH already produces some excess kurtosis through time-varying variance, but often not enough to match the data. By allowing conditionally $t$-distributed errors, Bollerslev separates two distinct sources of unconditional leptokurtosis — conditional heteroskedasticity and a conditionally leptokurtic error distribution — and lets the data apportion between them. The model is estimated by maximum likelihood and shown to describe foreign-exchange rates and stock-price indices better than the Gaussian GARCH.

Key Claims

• Two sources of fat tails. Volatility clustering (GARCH) generates leptokurtic unconditional returns even with normal innovations; but a conditionally heavy-tailed error adds a second, distinct channel. The GARCH-$t$ nests the Gaussian case and identifies which channel the data need.
• Standardized Student-$t$ innovations. The conditional distribution of $y_t$ given the past is a standardized $t$ with conditional mean $\mu_{t\mid t-1}$, conditional variance $h_{t\mid t-1}$ (the GARCH recursion of Bollerslev 1986), and degrees of freedom $\nu>2$; the variance is finite only for $\nu>2$, and tails get heavier as $\nu\to2$.
• Derivable as a variance mixture / subordinate process. The standardized $t$ arises as a continuous scale mixture of normals — equivalently by adding an unobservable multiplicative error to the conditional-variance equation — placing GARCH-$t$ within the scale-mixture-of-normals family and the subordinate-stochastic-process tradition (Clark 1973; Press 1967).
• Maximum-likelihood estimation. Via the prediction-error decomposition the log-likelihood is $\sum_t \log f_t(\varepsilon_t\mid\mathcal F_{t-1})$ with $f_t$ the standardized-$t$ density; standard ML inference applies to all parameters including $\nu$.
• Testing conditional normality is a boundary problem. Gaussianity corresponds to $\nu^{-1}=0$ (i.e., $\nu\to\infty$), a value on the boundary of the parameter space, so the usual $\chi^2$ asymptotics for the likelihood-ratio test do not apply directly.
• Empirical fit. Applied to a set of exchange rates and stock indices, the GARCH-$t$ improves descriptive validity over Gaussian GARCH, with estimated degrees of freedom low enough to indicate genuine conditional leptokurtosis beyond what volatility clustering alone explains.

Concepts Introduced or Extended

• GARCH and BEKK-GARCH
• Scale Mixture of Normals
• Heavy-Tailed Distributions
• Stylized Facts of Asset Returns

Entities Mentioned

• Tim Bollerslev

Quotes

"A simple extension of the ARCH model to allow for conditionally $t$-distributed errors is given. This development permits a distinction between conditional heteroskedasticity and a conditional leptokurtic distribution, either of which could account for the observed unconditional kurtosis in the data."

My Take

The lasting idea is the clean separation of two effects that a Gaussian GARCH conflates: time-varying variance and heavy conditional tails both fatten the unconditional distribution, and only by fitting a conditional $t$ can you tell how much of the observed kurtosis is dynamics versus distribution. That distinction became standard equipment — $t$ (and later skew-$t$ and GED) innovations are the default in applied volatility modeling, and the scale-mixture derivation is exactly the hook the wiki's scale-mixture and heavy-tailed stochastic-volatility pages build on for Bayesian estimation. The one technical wrinkle worth flagging is the boundary-testing issue for normality ($1/\nu=0$), which the paper notes and which recurs whenever a nuisance parameter sits on the edge of its space.