On Multivariate Time Series Model Selection Involving Many Candidate VAR Models

Xindong Zhao, Guoqi Qian


Vector autoregressive (VAR) models are important and useful for modelling multivariate time series. An appropriate VAR model is often required for such modelling for given data, for which several model selection criteria such as AIC, AICc, BIC and HQ are available. However, when the number of candidate models available for selection is extremely large, which is not uncommon in practice, performing an exhaustive VAR model selection using any of the above criteria would become computationally infeasible. To overcome this difficulty, we have developed a Markov chain Monte Carlo

method based on Gibbs sampler. It is shown that the developed method

identifies the optimal VAR model with high probability and

efficiency. To illustrate and verify the method, we also present a simulation study and an example on modelling the data of China's money supply and consumer price index (CPI).


Vector autoregressive (VAR) models, Gibbs sampler, Multivariate time series model selection

Full Text: