FAVAR

[Deepika]

Respected sir

i am trying to use FAVAR technique based on "MEASURING THE EFFECTS OF MONETARY POLICY: A FACTOR-AUGMENTED VECTOR AUTOREGRESSIVE (FAVAR) APPROACH", BEN S. BERNANKE & BOIVIN (2005). however when the draw starts for factors given factor loadings and VAR coefficients, then the program starts and it does not end giving me any result. how long should the program take. i am using RATS 9 Pro. i am attaching the data file and the program file for your reference. please help.

Regards
Deepika

[TomDoan]

You have data that can't be well-described by a FAVAR. X10 has one enormous spike. X25 is almost all dead flat with a couple of discrete changes. There are probably as least a dozen other series which can't be fit by this type of model.

[Deepika]

Respected sir

when i try changing the period of study starting from 2001m5 instead of 1996m4 the program does run. i still have to analyse the results. there is one more issue that under dlm command there is an option of limit. why is that taken to be 30 for the bernanke example. could you please explain? what should that be for my case?

Regards
Deepika

[TomDoan]

Without the LIMIT option, the BBE example can take something like 12 hours to run. I checked different values and found that 30 was more than enough (the filter generally converges in closer to 10-15).

[Deepika]

Respected Sir

Thanks for the reply. there is another issue that i have. the impulse responses that we get after running the program on FAVAR gives us the effect of a shock on the dependent variable to the independent variables in the system. how will the codes change if i am interested in giving a shock to the independent variables and analyse the effect on the dependent variable. thanks in advance for the effort that you and your team makes.

Regards
Deepika

[TomDoan]

They are all dependent variables; some just have a different role than others. What you're asking is not what this model is set up to provide.

[Deepika]

Respected Sir

what i mean to ask is that we have chosen some key variables in the model. can we not shock these key variables and analyse the response of the observable variable. we can do that in a standard VAR model. why cant we do that here??

Regards
Deepika

[TomDoan]

Because it's not a VAR. You can't do a VAR with 100+ variables or anything close to that. This model severely restricts the interactions among variables. In the BBE case, the only observable variable that can directly affect the others is the interest rate because it's the only one in the core VAR with the unobservable factors. The other variables get information from the core VAR, but don't put information into it.

[Deepika]

Respected Sir

After having gone through some literature on FAVAR, it seems that the way Bernanke et al have set up the model may be one of the way of doing it. For example the paper by Haroon Mumtaz & Paolo Surico, 2009. "The Transmission of International Shocks: A Factor-Augmented VAR Approach," Journal of Money, Credit and Banking, Blackwell Publishing, vol. 41(s1), pages 71-100, 02, follows a relatively different approach. could you please tell me as to how the codes would be different. what i could understand from the paper was that they extracted factors using principal component analysis first and then rely on interplay of factor loadings to give a more meaningful insight to the results.

Regards
Deepika

[TomDoan]

I don't think you're reading that correctly. It's the same basic idea as BBE, but with a somewhat larger set of factors and a structural contemporaneous model. That doesn't change the fact that you can't analyze the effects of shocks to series that aren't part of the core VAR.

[macrograd]

In response to reading this thread, is there code available for replicating the BBE FAVAR using the "two-step" principal components approach? Am I correct that the replication file only does the one-step likelihood approach?

It looks as if the Martin, Hurn, Harris example MHHP564 is very close and could be modified to do this but I'm wondering how to extract the response functions for any element of Xt (any of the 120 series in BBE).

Thanks for any answers and help you could give.

[Deepika]

Respected Sir

There are some more issues that are not clear with the program:
1) How do you define the factors in this program. BBE paper says that the way we define the slow moving variables helps us to define the factors. but isn't there a specific way of doing it??? so if i have 3 factors then how do i know which one represents the real activity or international prices or credit situation in the economy and so on.
2) If i were to forecast a macroeconomic variable using FAVAR, then is it possible to do that as an extension of the existing program codes or do we have some other set of codes that we use here??


Regards
Deepika

[TomDoan]

In response to reading this thread, is there code available for replicating the BBE FAVAR using the "two-step" principal components approach? Am I correct that the replication file only does the one-step likelihood approach?

It looks as if the Martin, Hurn, Harris example MHHP564 is very close and could be modified to do this but I'm wondering how to extract the response functions for any element of Xt (any of the 120 series in BBE).

Thanks for any answers and help you could give.

The example in MHH is for a simpler model where all the observables are in the "Y" category. (While they don't indicate it in the book, they are able to handle 30 variables by assuming the gamma matrix is diagonal. We got that out of the programs that they provided). BBE's "two-step" procedure actually requires more than that, because you have to transform the PC's into the space that's orthogonal to the Y's in the VAR.

[TomDoan]

Respected Sir

There are some more issues that are not clear with the program:
1) How do you define the factors in this program. BBE paper says that the way we define the slow moving variables helps us to define the factors. but isn't there a specific way of doing it??? so if i have 3 factors then how do i know which one represents the real activity or international prices or credit situation in the economy and so on.

You don't (you will note that they never say what the factors are). They are to soak up as much as possible of the common movements of the other 119 series so the monetary shock is better isolated. As with typical factor analysis, you can attempt to infer what a factor represents by how much of various variables it explains. Note that the RATS program includes this, but the BBE paper only includes the percentage due to the monetary shock. However, there's very little attempt in the BBE model to try to differentiate the shocks in any meaningful way---there are minimal restrictions to make them statistically identified.

2) If i were to forecast a macroeconomic variable using FAVAR, then is it possible to do that as an extension of the existing program codes or do we have some other set of codes that we use here??

The only series that it's really designed to forecast at all are the "Y"'s (and it's not clear that it's "designed" to forecast even those). This is similar to the Stock and Watson model where the vast majority of the variables exist to improve the factors, but the model is designed to forecast only a few core variables.

[Deepika]

Respected Sir

Sorry to ask this but the Stock and Watson example that you are talking about, do we have a program example for that??

regards
Deepika