Not all long‐memory estimators are born equal: The case of nonstationary functional time series

Abstract

We study a functional version of fractionally integrated time series that covers the nonstationary case when the memory parameter d is above 0.5. We project time series, with varying levels of nonstationarity, onto a finite‐dimensional subspace. We obtain the eigenvalues and eigenfunctions that span a sample version of the dominant subspace through dynamic functional principal component analysis of the sample long‐run covariance functions. Within the context of functional autoregressive fractionally integrated moving average models, we evaluate and compare finite‐sample bias and mean‐squared error among some time‐ and frequency‐domain Hurst exponent estimators via Monte Carlo simulations. We apply the estimators to Canadian female and male life‐table death counts.