Simultaneous Detection of Different Types of Shocks in Normal Matrix Linear Dynamic Models

Abstract

In this work, the process of detecting various types of shocks in a Matrix Linear Dynamic Normal Model (MLDNM) is framed as a problem of Bayesian model comparison. This approach allows for analyzing the existence of a wide range of atypical behaviors in the evolution of multivariate time series (such as isolated outliers, level changes, slope changes, seasonal pattern shifts, etc.) both simultaneously and sequentially. Based on the detected shock type, interventions can be made. The proposed framework extends the algorithm for automatic monitoring and intervention introduced by Gargallo and Salvador (2002c) for univariate time series analysis. As an illustration, the procedure is applied to the analysis of the evolution of mortgages and deposits in the banking system in Aragon