An algorithm for robust fitting of autoregressive models

whose unique solution φ̂1, . . . , φ̂p and σ̂2 forms the well-known YW estimator that is is asymptotically efficient in the context of a Gaussian AR series. Nevertheless, the YW estimator loses its asymptotic efficiency under a non-Gaussian distributional assumption; see e.g. Sengupta and Kay (1989). In what follows, we describe a simple estimation algorithm for AR model fitting; it is not a fast algorithm but it is promising in improving the finite-sample accuracy of the YW estimators when outliers are present. The new algorithm exemplifies robustness against outliers, and in particular against clusters of (two or more) outliers. ∗Department of Mathematics, and Economics, Univ. of California at San Diego, La Jolla, CA 92093-0112, USA; tel.: (858) 534-5861, fax: (858) 534-5273, e-mail: [email protected]. Research partially supported by NSF grant SES-04-18136.