The RATS Software Forum

Dear Tom,

Could I have a reference about Markov-Switching SVAR estimation & identification with RATS? Or any suggestion though!

Best,

It depends upon what the "S" in SVAR means. If you have a just-identified structural model, you can just do the standard MS-VAR and do the structural part separately, as is typically done with SVAR's. If it's over-identified, it's a lot more complicated.

My suggestion for anyone exploring MS models is to first read the "How to Switch if You Must" article in the January newsletter.

TomDoan wrote:It depends upon what the "S" in SVAR means. If you have a just-identified structural model, you can just do the standard MS-VAR and do the structural part separately, as is typically done with SVAR's.

Dear Tom,

Thank you for your reply.

The SVAR analysis builds upon how we restrict A and B in A*u(t) = B*v(t). In your words, is it possible if I estimate the VAR-X (simply let some exogenous vars in the "det" supplement), getting the %sigma and other suitable model components,... The next step is to follow the usual identification schemes as such A, B or A-B as long as I have a just-identified restriction. Right? That is the way SVARX works?

Best,

There's no real difference in estimation between a VAR and a VAR-X. Both rely on the properties of multivariate regressions with the same explanatory variables in each equation. That (in general) is also true if you add a structural contemporaneous model.

Given a sample, an SVAR (or SVAR-X) can typically be estimated in two stages by estimating the lag coefficients and the concentrated covariance matrix of residuals from it, and then estimating the coefficients of the structural model using the covariance matrix only. That's true whether or not the structural model is just-identified or over-identified because the maximum likelihood estimates of the lag coefficients are the same regardless of covariance matrix. If the sample is not known (because you have a model with multiple regimes), it no longer works to do "two stage" estimates for over-identified models as the log likelihoods of the regimes are affected by the overidentifying restrictions. It does still work for just identified structural models as you can estimate the switching model estimating only regime-dependent covariance matrices and then can estimate the structural models based upon them as the existence of the structural model doesn't change the maximizing log likelihood of the regimes.

Dear Tom,

Can you help me with the code to get the results of the model Markov switching (identified) structural GARCH‐in‐Mean VAR proposed by Apostolos Serletis and Libo Xu (2019). Markov Switching Oil Price Uncertainty. Oxford Bulletin of Economics and Statistics, https://doi.org/10.1111/obes.12300

Best regards

I don't think so. You might want to read two articles from the January 2018 newsletter. Not just the "How to Switch if You Must", but also the "Markov Switching GARCH Models".