Serial dependence in ARCH-models as measured by tail dependence coefficients (Q2271709)

The tail dependence of an ARCH(1) process \(X_i\) is investigated in terms of the lower tail dependence coefficient \[ \lambda_{X| Y}=\lim_{\alpha\to 0}P\{X\leq q_X(\alpha)\;| \;Y\leq q_Y(\alpha)\} \] (\(q_X\) being the quantile function of \(X\)) and its generalizations. The coefficient \(\lambda_{X_{n+p}| X_n}\) for stationary \(X_i\) is evaluated as an integral of the CDF and PDF of innovations. Asymptotics of transition probabilities and quantiles are investigated for nonstationary \(X_i\). As a result it is shown that for ARCH(1) processes with fixed nonrandom initial value and symmetrically distributed innovations \(\lambda_{X_{n+p}| X_n}=0\).