Keywords: ARCH, ARMA Model for a Squared GARCH, Autocorrelation, Autocovariance of a Squared GARCH, Ergodicity, GARCH(p,q), IGARCH, Infinite ARCH, Lyapounov's Coefficient, Kurtosis, Moments, Prediction of the squares, Riskmetrics, Shock Persistence, Strict and second-order stationarity, Strong and Semi-Strong GARCH, Top Lyapounov Exponent, Volatility, Volatility clustering.
Description: ARCH (AutoRegressive Conditionally Heteroskedastic) models were introduced by Engle (1982) and their GARCH (Generalized ARCH) extension is due to Bollerslev (1986). In these models, the key concept is the conditional variance, that is the variance conditional to the past. In the classical GARCH models, the conditional variance is expressed as a linear function of the squared past values of the series. This particular specification is able to capture the main stylized facts characterizing financial series, as described in Chapter 1. In the same time, it is simple enough to allow for a complete study of the solutions. The 'linear' structure of these models can be put forward through several representations that will be studied in this chapter. In this chapter, we first present definitions and representations of GARCH models. Then, we establish the strict and second-order stationarity conditions. Starting with the first-order GARCH model, for which the proofs are easier and the results are more explicit, we extend the study to the general case. We also study the so-called infinite ARCH models, which allow for a slower decay of squared-returns autocorrelations. Then, we consider the existence of moments and the properties of the autocorrelation structure. We conclude this chapter by examining forecasting issues.
R code
used to plot the stationarity regions in Figures 8 and 10 , Mathematica code
used for the correlograms of Figure 11, and
for the correlograms of Figure 12. Fortran program
for computing ordinary and partial autocorrelations (of the squares) of an ARMA-GARCH (in F90 with Absoft's compiler) +
executable program +
data file for the program (example of Figure 12)