'lvfxvvlrq Ri 6hvvlrq 0rgho 6hohfwlrq Dqg (ydoxdwlrq

selecting a IUDFWLRQDO H[SRQHQWLDO )(;3 model for a Gaussian long-memory time series. A regression approach is used to obtain the least squares estimator of d of the memory parameter, and mean squared error is minimized. The paper overlooks the fact that in the case of regression models, there is a relationship between $NDLNH¶V LQIRUPDWLRQ FULWHULRQ $,& and 0DOORZV¶ / & , which can also be used in this problem. For this, see, e.g., Bozdogan and Ramirez (1986). It is remarkable that / & is applied only in the case of linear regression models whereas $,& is used in more general modeling situations. Based on this relationship one can also establish the asymptotic unbiasedness, up to irrelevant additive and multiplicative constants, of a local $,& criterion for the PHDQ VTXDUHG HUURU 06( of a fractional estimator of the memory parameter. A number of questions come to mind when reading this paper. These are: • How are the appropriate choices of the number of frequencies, m, carried out? • What are the consistency properties of the estimator of d? • Can we consider using kernel type estimators in this problem?