The problem of statistical inference for the mean of a time series with possibly heavy tails is considered Wefirst show that the self-normalized sample mean has a well-defined asymptotic distribution Subsampling theory is then used ot develop asymptotically correct confidence intervals for dthe mean without konowledge (or explicit estimation) either of hte dependence characteristics or of the tail indux Using a symmetrization technique we also construct a distribution astimator that combines robusteness and accuracy:it is higher-order accurate in the regular dase which remaining consistent in the heavy tailed ased Some finitesample simuations confirm the precticality of ht proposed methods.
展开▼
