The RATS Software Forum

Hi

I am currently trying to study the method of GVAR and regime switching models. My main research interest is to study the regime-dependent inter-linkages between international capital flows and domestic financial variables in a global setting.

Based on my understanding, GVAR is implemented by estimating individual country's VARX* model by including country-specific variables X* (computed using trade weight), and afterwards each country's VARX* is combined into global model. That being said, is it possible if I estimate Threshold VARX* instead for each country in the first stage and combine them into global model as per normal?

(What confuses me is if I use endogenous domestic variable as threshold variable, e.g. domestic financial market risk, then it means for each country, they may have different state dynamic. For instance, it is possible for country 1 to be in state 1 while country 2 in state 2 within the same period. Then, is it viable to combine them into global model? Alternatively, should I just use global risk indicator as threshold variable so that all countries will have same state dynamic?)

I have actually read an ECB working paper about 'Regime-Switching Global Vector Autoregressive Models' by Binder and Gross(2013), but the code is not accessible, and I don't really know how to implement it.

Related Posts and Reference:
https://estima.com/forum/viewtopic.php? ... 1608#p1608 (GVAR)
https://estima.com/forum/viewtopic.php?f=8&t=2003 (TVAR)
https://www.ecb.europa.eu/pub/pdf/scpwp ... 4192419033 (Binder and Gross(2013))

The paper you're citing isn't using a threshold model---they're using Markov switching (thus hidden, rather than observable regimes), and yes, they allow the different countries to be in different regimes. With 18 countries, that means there are 2^18=262144 possibilities at each entry. How they can estimate that in any reasonable amount of time is beyond me.

Dear TomDoan,
Thank you so much for your reply.
I have seen the other posts in which you provided the replication codes for TVAR and GVAR. I wonder if it is possible to estimate each country's TVAR (with inclusion of X*, the trade weighted country-specific variables) following your codes, and then combine the individual TVARs of each state into global model using selection matrix (as described in pg 36 of the GVAR_Slides.pdf). Put differently, let say I use global risk indicator as threshold variable, and assume there are normal and high-risk regimes. For each regime, after getting the estimated coefficients, I should be able to combine all countries' individual TVAR into a global model right? My ultimate objective is to get GVAR representations for both normal and high-risk regimes, and then conduct impulse response analysis on the shock of one country's variable on other country's variables in each regime under such global setting. But I am not sure if this is viable or can it be implemented using the TVAR and GVAR codes that you provided.

Sorry if I sound clueless since I am still on early research stage and I really need some advice. Thanks!

https://fe61f33a-a-62cb3a1a-s-sites.goo ... edirects=0 (GVAR_Slides.pdf)

That's really outside the scope of technical support. You should probably address those questions to people who are working in the area (such as Pesaran). Again, however, I would point out that you seem to be describing a threshold model rather than a MS model, so you probably want to see whether there's any literature on threshold GVAR's.

Dear TomDoan,
Thanks for your advice. I am aware that the paper I cite is MS model, since I couldn't find GVAR paper that incorporate other types of regime-switching model. But I actually prefer TVAR due to its simplicity in inferring regime using a particular variable.
I should definitely check if what I intend to do is theoretically viable. That aside, I am also curious about the technical issue, as whether it could be implemented using RATS codes.

Estimating a separate TVAR for each country isn't especially difficult---see the Balke replication. One potential problem is that you need a certain number of observations in each regime (at least equal to the number of regressors) which might or might not be a problem depending upon how you are trying to separate regimes. The paper you cited was aggregating the separate country information based upon differing regimes, and that step, at least, doesn't depend upon how the regimes are identified (whether they're observable or latent). The one major drawback with your proposal is that you can't do IRF's in a threshold model one regime at a time---the regime switch is an integral part of the dynamics. Again, you should study the Balke example. You would also probably benefit a great deal from the Switching Models/Structural Breaks e-course.

The one major drawback with your proposal is that you can't do IRF's in a threshold model one regime at a time---the regime switch is an integral part of the dynamics.

What you mean is the regime switch itself is endogenous that is why it can't be achieved? But we can still get IRs conditional on regime through simulation as in Balke's paper right?
Sorry for so many questions...

I think you're misinterpreting what that's doing in the Balke paper. IRF's in a non-linear model depend upon a lot of things (which is why they require simulation/bootstrapping), but in particular, are different depending upon the initial conditions. What Balke did was to average across the initial conditions associated with each regime.