AN EM ALGORITHM FOR CONDITIONALLY HETEROSCEDASTIC FACTOR MODELS

This article discusses the application of the EM algorithm to factor m odels with dynamic heteroscedasticity in the common factors. It demons trates that the EM algorithm reduces the computational burden so much that researchers can estimate such models with many series. Two empiri cal applications with 11 and 266 stock returns are presented, confirmi ng that the EM algorithm yields significant speed gains and that it ma kes unnecessary the computation of good initial values. Near the optim um, however, it slows down significantly. Then, the best practical str ategy is to switch to a first-derivative-based method.