GIBBSVARBUILD—Gibbs sampling for Bayesian VAR
Posted: Thu Mar 21, 2024 10:47 am
GIBBSVARBUILD.RPF does Gibbs sampling on a Bayesian VAR, using the @BVARBuildPriorMN procedure to generate the mean and precision matrices for the prior. This is a modernized version of the GIBBSVAR.RPF example.
Unlike the calculations done using ESTIMATE, this treats the whole VAR as a unit rather than one equation at a time. This can be applied successfully to VAR's of modest size. This example has 4 variables x 4 lags (plus constant), so it has a total of 68 regression parameters. This will probably run in a reasonable time for up to perhaps 500 total parameters. The sticking point is that a full-system draw for the VAR coefficients requires taking a factor of a # of coefficients x # of coefficients matrix with no helpful structure to reduce the calculation time.
Detailed Description
Unlike the calculations done using ESTIMATE, this treats the whole VAR as a unit rather than one equation at a time. This can be applied successfully to VAR's of modest size. This example has 4 variables x 4 lags (plus constant), so it has a total of 68 regression parameters. This will probably run in a reasonable time for up to perhaps 500 total parameters. The sticking point is that a full-system draw for the VAR coefficients requires taking a factor of a # of coefficients x # of coefficients matrix with no helpful structure to reduce the calculation time.
Detailed Description