Topic: Time series: Nonparametrics and extremes Title: Long memory parameter estimation: Signal or noise Authors:  Grace Yap - The University of Nottingham Malaysia Campus (Malaysia) [presenting]
Wen Cheong Chin - Multimedia University (Malaysia)
Abstract: Long-memory parameter estimation using bias-reduced log-periodogram regression (BRLP) $\hat{d}_r$ is proven efficient as it eliminates the first and higher order of biases of the log-periodogram model. Nonetheless, its performance relies largely on the frequency bandwidth $m$ and the order of estimation $r$. Literature suggests a data-dependent plug-in method for selecting the frequency bandwidth that minimizes the asymptotic mean-squared error (MSE). However, this choice of $m$ significantly increases the MSE's over the finite sample minimum MSEs due to the non-parametric estimation problem in the unknown term within the plug-in method. In a long-memory time series with mild short range contamination, a simple approach to determine the bandwidth size is suggested based on the spectral analysis. Monte Carlo simulation results for stationary ARFIMA $(1,d,0)$ and ARFIMA $(0,d,1)$ processes show that with the proper order of estimation, the proposed bandwidth selection performs better than that of the MSE optimal choice.