Computational aspects of maximum likelihood of autoregressive fractionally integrated moving average models

Computational aspects of likelihood-based estimation of univariate ARFIMA (p,d,q) models are addressed. Particular issues are the numerically stable evaluation of the autocovariances and efficient handling of the variance matrix which has dimension equal to the sample size. It is shown how efficient computation and simulation are feasible, even for large samples. Implementation of analytical bias corrections in ARFIMA regression models is also discussed. © 2002 Elsevier Science B.V. All rights reserved.