ROBUST METHODS FOR RECURSIVE AUTOREGRESSIVE MOVING AVERAGE ESTIMATION

In this paper we consider the problem of fitting, recursively in time, an autoregressive moving average (ARMA) model to time series data whe re outliers may be present. Although many recursive estimation procedu res are available for fitting ARMA models, they are based on the recur sive least squares algorithm which is known to be badly affected by ad ditive outliers. To minimize the impact of these sort of outliers we i nvestigate some robustified versions of the recursive maximum likeliho od procedure. The problem of incorporating robust scale recursion is d iscussed and asymptotic properties of the procedures are examined. We also present results from simulation studies which compare adaptive ve rsions of these procedures in the non-stationary case.