Second order properties of distribution tails and estimation of tail exponents in random difference equations (Q626302)

Stationary solutions \(X_n\) to a random recurrence (difference) equation (RDE) are considered defined by \(X_n=A_nX_{n-1}+B_n\), where \(A_n\) and \(B_n\) are i.i.d. random variables. Estimation of tail index \(\alpha\) for \(X_n\) marginal distribution is discussed. It is noted that the usual Hill estimate is biased since, even with a huge number of data points, there are significant deviations from the Pareto tail of \(X_n\) in practice. The authors consider models which satisfy the condition of second order regular variation for RDE solutions. In this case, an estimate for \(\alpha\) is constructed by regressing the logarithm of empirical tail distribution on \((1,\log x,1/x)\). Behaviour of the estimate is investigated via simulations. ARCH models and multiplicative cascades are considered as examples. A multivariate extension is discussed also.