Multiple Time Series Modeling II.

Abstract : This paper defines the problem of time series modeling as model identification (determining the predictor variables) and parameter identification (estimating the prediction filter and the prediction error covariance matrix). Various auto-regression and cross-regression representations are defined for a stationary multiple time series. The role of basic regression and latent value algorithms is discussed. It is suggested that principal component analysis of spectral density matrices may not be useful in practice, whereas autoregressive methods are. The problem of defining an index time series is discussed; an approach is described in terms of the notion of predictable components. (Author)