Canonical Correlation in Multivariate Time Series Analysis With an Application to One-Year-Ahead and Multiyear-Ahead Macroeconomic Forecasting

Abstract

A simple one-period-ahead and multiperiod-ahead prediction procedure for multivariate time series is suggested, based on the canonical correlation technique. The prediction procedure is direct in the sense that no lag orders and parameters have to be estimated first, as in the usual ARMAX or VAR parameterizations of multivariate stationary stochastic processes. A best (in the mean squared error sense) predictor can be obtained directly using singular-value decompositions of covariance matrices. The procedure is used to forecast one-year-ahead and multiyear-ahead national growth rates of 14 countries for the years 1974–1984.