steve-the-bayesian / BOOM

A C++ library for Bayesian modeling, mainly through Markov chain Monte Carlo, but with a few other methods supported. BOOM = "Bayesian Object Oriented Modeling". It is also the sound your computer makes when it crashes.
GNU Lesser General Public License v2.1
36 stars 15 forks source link

mbsts suggested improvements #53

Open steve-the-bayesian opened 4 years ago

steve-the-bayesian commented 4 years ago

I thought I would bring the following to your attention:

  1. It would be nice if we could specify alternative models for the shared factors, other than the random walk implied by AddSharedLocalLevel. In particular, a stationary AR(1) process for the factors would be very useful. You are no doubt working on this already.

  2. With univariate BSTS models, the fitted model objects can get very large when the number of MCMC iterations is large. I therefore often use the model.options parameter to omit results that I do not need (e.g. model.options = BstsOptions(save.state.contributions=FALSE, save.prediction.errors=FALSE)). This functionality is likely to be even more important in the multivariate case.

  3. Missing values seem to be permitted only in a subset of the series at each point in time. If all series have missing values at a point in time (e.g. on a public holiday), mbsts throws this error: Caught exception with the following error message: object of length 6 does not conform with the number of columns (0)