Time Series, 2nd Edition

364 LATENT FACTOR MODELS

particular case of the hierarchical structural constraint introduced in Sec-

tion 11.2.4. Another way to resolve the identifiability problem is to assume

that the length of the vector B is one, that is,

P

r

i=1

b

2

i

= 1. Although this

is relatively straightforward to implement computationally for the case of

one factor, its extension to more than one factor in a Bayesian framework

is far from trivial. We return to this issue in Section 11.2.4.

Further, we assume that the other factor loadings in the vector B are un-

constrained. In particular, in a Bayesian framework we have to assume a

probability model for the factor loadings. Here ...