MODEL-SPECIFICATION TESTS FOR BALANCED REPRESENTATION STATE-SPACE MODELS

Balanced representation state space models of the form developed by Aoki (1990) have several inherent advantages relative to ARMA models in the specification of multivariate time series. These include an orderly specification search over a small number of dimensions and, as demonstrated below, a choice of coordinate systems for the states that makes sequential model order tests uncorrelated, thus preserving their confidence levers. Using two representative data sets as the basis for a Monte Carlo experiment, four model specification procedures are evaluated. The results show promise for two of the methods which are based on asymptotic likelihood ratio tests.