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
need priors specified in terms of discount factors #3
The current default priors in bsts assume innovation variances are "small" relative to the sample variance. This is sensitive to the length of the series (what is "small" for a short series might not be small for a long series), and arbitrary.
For state models with constant model matrices, one can find the steady state variance for the model (without actually using the data y). This can be done analytically for the local level model, but things get hard after that, so it needs to be done in real time, and it probably requires access to the residual variance sigma^2.
Subtracting RQR from the steady state variance gives a variance that can be interpreted in terms of discount factors.
The current default priors in bsts assume innovation variances are "small" relative to the sample variance. This is sensitive to the length of the series (what is "small" for a short series might not be small for a long series), and arbitrary.
For state models with constant model matrices, one can find the steady state variance for the model (without actually using the data y). This can be done analytically for the local level model, but things get hard after that, so it needs to be done in real time, and it probably requires access to the residual variance sigma^2.
Subtracting RQR from the steady state variance gives a variance that can be interpreted in terms of discount factors.