djpedregal
commented
3 years ago Thank you very much for your interest in SSpace. Unfortunately SSpace does not handle multivariate non-Gaussian cases, so the answer to your first question is that it cannot be directly implemented. There may be, however some sort of Gaussian approximation. Check a references by Harvey and Esther Ruiz about Stochastic volatility approximated by linear Gaussian models, I am not sure it would do.
Regarding the second question, the help about the compulsory ‘H’ is in terms of code, it means that you should not remove the H from the inputs to the SampleNONGAUSS template because SSpace would crash. In the case you comment you are right, H is not used and SSpace uses H = 0 internally for the case you comment. It may be different from zero for other non-Gaussian models, heavy T-tailed or Stochastic Volatility models. See Durbin and Koopman (2012). You may see an example in demo7: the user does not specify H and SSpace assign a value of 0 to it internally.
I hope this helps. Best wishes. Diego.
El 30 dic 2020, a las 10:21, Peter_Xu notifications@github.com<mailto:notifications@github.com> escribió:
Dear Dr. Pedregal,
I have an ss model that y=(y1, y2) is two-dimensional where y1 follows a normal distribution and y2 follows a binomial distribution (n is known). They are conditionally independent given the latent state. I am wondering whether the sspace toolbox can handle this kind of question. If it does, how can I write model.dist and model.param? I guess model.dist = ['Gaussian' 'Binary']? (but 'Gaussian' is not in the distribution list).
My second question is that in your document, in the non-gaussian model, yt ~ p(yt| theta_t)+ Dt, and theta_t = Z_t alpha_t. So that there is no H_t matrix involved. Then in SampleNONGAUSS.m, it says (line 17), 'Beware that the 'H' input to this function, i.e. the observational noise variance, is compulsory.' I'm not sure why do we need H, and what is this H_t referring to in the non-gaussian model.
Thank you!!
Regards Peter
— You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHubhttps://github.com/djpedregal/SSpace/issues/2, or unsubscribehttps://github.com/notifications/unsubscribe-auth/AG3UEQ74BT7TGALE52BR2H3SXLWH7ANCNFSM4VOD2W6Q.
zjph602xtc
Dear Dr. Pedregal,
I have an ss model that y=(y1, y2) is two-dimensional where y1 follows a normal distribution and y2 follows a binomial distribution (n is known). They are conditionally independent given the latent state. I am wondering whether the sspace toolbox can handle this kind of question. If it does, how can I write
model.distandmodel.param? I guessmodel.dist = [Gaussian Binomial]? (but 'Gaussian' is not in the distribution list and this seems to return an error).My second question is that in your document, in the non-gaussian model, yt ~ p(yt| theta_t)+ Dt, and theta_t = Z_t alpha_t. So that there is no H_t matrix involved. Then in SampleNONGAUSS.m, it says (line 17), 'Beware that the 'H' input to this function, i.e. the observational noise variance, is compulsory.' I'm not sure why do we need H, and what is this H_t referring to in the non-gaussian model.
Thank you!!
Regards Peter