Performance of variance function estimators for autoregressive time series of order one: asymptotic normality and numerical study

We study performance of several conditional variance estimators for an autoregressive time series which include local linear smoothers with various bandwidths, local likelihood and difference-based estimators. In the theoretical part, asymptotic normality of the local linear estimator of variance with no mixing assumptions imposed on the underlying process is proved. Moreover, numerical examples performed reveal that a two-stage local linear smoother with a bandwidth, proposed by Ruppert, Sheather and Wand, used to estimate the regression function and a simple rule of thumb bandwidth for variance estimation performs best for variances without much structure, whereas the bandwidth considered by Fan and Yao works very well for much more variable variances.