A computational method for estimating continuum factor models
Let \(z_t\) be a stationary k-dimensional vector time series following the p-th order vector autoregressive model. A linear transformation of the vector process is proposed to reduce the number of involved parameters. This leads to the so-called ``continuum factor model \(z_t=v+ B_1f_{t-1}+\dots+B_pf_{t-p}+\epsilon_t\), where \(z_t\) is determined from the set of \(r<k\) common hidden factors \(f_t=cz_t\) carrying the important forecasting information. The paper focuses on computational methods for estimating the factors \(f_{it}=c^Tz_t\), \(i=1,\dots,r\), generalizing Stone and Brook's approach. Weights \(c_i\) can be found by solving the first and the \(m\)-th canonical problems. The authors show the possibility to transform the \(m\)-th canonical problem to the fist one, which is solved using Lagrange multipliers and the Gauss-Newton procedure. A method to get good initial vectors for the Gauss-Newton procedure is proposed, too. Properties of the computational method are discussed in detail. It is proved that \(c_i\) are consistently estimated when the sample path covariance is used.
This page was built for publication: A computational method for estimating continuum factor models
Report a bug (only for logged in users!)Click here to report a bug for this page (MaRDI item Q1389255)
