Structural Breaks and VAR Modeling with Marginal Likelihoods

Wolfgang Polasek - University of Basel

Structural breaks can be tested with various classical methods (like the Chow test or the likelihood ratio test). For time series models decisions can be made with the help of an information criterion (AIC, BIC, etc.), especially if the lag length is unknown as well. Using the Bayesian concept of marginal likelihoods which is also an intermediate step to derive the Bayes factors for the posterior odds (or Bayes tests) of hypotheses testing. Using the marginal likelihood identity of Chib (1995) we derive the marginal likelihoods for AR andV AR models with and without breaks. It is shown with an simulated example that the marginal likelihood criterion has a better frequency performance than the classical test statistics. A macro-economic example involving Swiss consumptiona nd GNP shows how this approach can be used to explore and analyze multivariate regime shifts. In a final outlook it is shown how this approach can be extendedt o error correction and cointegration models where Gibbs sampling methods have to be used.

Keywords: information criterion, marginal likelihood, regime shifts, order estimation.

Scheduled for Session 3.2 Time Series - II