srlanalytics
commented
6 years ago Could do. I had focused on Bayesian estimation because:
(1) It is the hardest one to code up correctly
(2) It is typically the most powerful for forecasting applications due to the ability to include prior information
(3) There are other packages that do ML dynamic factor models using Watson and Engle's version of the EM algorithm (which vastly superior to numerical optimization of the likelihood function)... this is how we do it too for method = "ML". There are no packages that I'm aware of (or can find easily on google) that do Bayesian estimation of factor models using Durban and Koopman's disturbance smoother. The issue is that factors are not identified; Durbin and Koopman (2012) write that:
Implementing the two steps of the MCMC is not as straightforward as suggested so far...
Though there are other approaches, I worked out a reliable methodology for that bit on my own.
(4) Some people (like me) think that a Bayesian approach to statistics is the "right" way to view the world... the frequentest idea of "true" parameters is not very satisfying.
BDFM is the jargon used in the literature for Bayesian dynamic factor models and is thus appealing as it means something on it's own. However, if you think a broader approach would be easier to sell that's probably a better way to go. Let me know what you think...
christophsax
Now, since we cover other methods than Bayesian as well, why not advertise it a bit more broadly?
E.g. call it dfm or dfmr, both names are available. The latter would have the advantage that it is easy to search for your package on the internet.