Information complexity criteria for detecting influential observations in dynamic multivariate linear models using the genetic algorithm

We develop a new information theoretic approach for detecting influential observations in dynamic linear models of multivariate time series known as vector autoregressions (VARs). Our approach consists of two stages. In the first, we use a Genetic Algorithm (GA) with Bozdoganʹs informational complexity (ICOMP) criterion as the fitness function to select a near optimal subset VAR model. In the second stage, we use ICOMP with case-deletion on the subset VAR chosen by the GA to detect influential observations. Our approach yields an intuitive, practical and rigorous two-dimensional graphical representation of influential observations in multivariate time series data that accounts for both lack-of-fit and model complexity in one criterion function. We demonstrate our approach on multivariate macroeconomic time series data.