Robust model order selection for ARMA models based on the bounded innovation propagation τ-estimator

A crucial task when fitting an ARMA model to real-world data is the selection of the autoregressive and moving-average orders. In real-world applications, the data may contain measurement artifacts or outliers (aberrant observations). Robust model order selection aims at finding a suitable statistical model to describe the majority of the data while preventing outliers or other contaminants from having overriding influence on the final conclusions. Three new approaches for robustly selecting the ARMA model orders based on the bounded innovation propagation (BIP) τ-estimator are presented. These are compared via Monte Carlo simulations to existing robust and non-robust criteria. A real-data application of ARMA modeling for artifact removal in intracranial pressure signals is provided.