A Computational Method for Estimating Continuum Factor Models
Computational Statistics, Vol. 12, No.4 (1997)
Posted: 2 Mar 1998
This paper focuses on the computational aspects of a forecasting method for multiple time series, the so called continuum factor models proposed by Sjostedt (1996). Within the vector autoregressive framework, linear transformations of the vector process are considered, revealing possible simplifying structures, as hidden factors catching the important forecasting information. The factors are solutions to nonlinear constrained optimization problems. These optimization problems have a special structure that the solution procedure takes advantage of. The main ideas are to transform the optimization problem to problems with only one constraint, using the Lagrange equations and the Gauss-Newton method. Methods to get initial vectors to the Gauss-Newton method are also proposed. Properties of the computational method are analyzed and discussed. It is also proved that the continuum factors are consistently estimated when sample covariance matrices are used.
JEL Classification: C22, C32, C63
Suggested Citation: Suggested Citation