FAVAR add-in

[Bens]

Hey everyone,

I am also using this great add-in to do a favar in eviews. I am trying to update the analyses of Bernanke et al (2005) and I am just wondering if anyone knows, why the program also standardizes the federal funds rate? I get that the variables for the principle component analysis have to be standardized but I dont think that this is should be the case for the ffr. Or did I just didnt get the concept?

Furthermore it seems like since 2001 there is a unit root in the federal funds rate, which persists even when we use the shadowrate instead. Does anyone know how to handle something like this in the VAR?

Help would be greatly appreciated

Ben

[dakila]

The fed fund rate is observable factor. Other factors are unobservable. Therefore FFR should be compatible with other factors. For example, FFR is used with the factor rotation analysis.
Even though FFR is I(1) variable, you cannot difference FFR. Because it is impulse variable. Shock to FFR should decay slowly.

[Bens]

Hey dakila,

thanks for your fast response. Your add-in is great and it saved us a lot of time. Regarding the scaling of the FFR: Thanks the clarification. I didnt know that but intuitively it makes sense now.

Regarding the unit root in the ffr: The point is that when using our data (which is an updated version of the BBE data) and doing a favar with the ffr as exogenous variable, the impulse response of the ffr to an ffr shock does not die out, when we restrict the sample to start in 2009M01 and end in 2017M11. Do you know any way around that problem? We suspect that the unit root in our exogenous variable changes our results, since we see the reappearance of the price puzzle over the course of the crisis. Also in the original BBE dataset some variables are not stationary (most notably all variables concerning interest rates). Is that no problem as long as the factors are stationary?

Furthermore: is it possible to scale the IRFS in your great add-in? We wanted to have a 25 basispoints increase in the ffr in order to compare it to the BBE results. I know there is a very nice eviews add-in for scaling IRFs but does it work with your favar add-in?

Thank you for your help. I really! appreciate it

Best,
Ben

[dakila]

Hi Ben,

The FAVAR add-in is updated. Now it includes an option to scale the IRF. Please update it.

Is that no problem as long as the factors are stationary?

Yes, it is no problem.
Many researcher estimate VAR model in level variable (not first difference) even though variables are I(1).
For example, Chris Sims, father of VAR model estimate it in level. Of course shock to FFR does not die out.
Do not worry about it. You also should look at Variance Decomposition. According to VD analysis monetary policy shock usually explain small fraction of real variable (for example, real GDP).

Please remember that in order to scale a 25 basis point, you should divide it by standard deviation of FFR. For example: 0.25/3.1987=0.07816

[Bens]

Hey dakila,

thanks so much for your help. And many thanks for the update. The scaling possibility is awesome and also the fact that one can now safe the impulse responses to a matrix. Regarding the add-in I have two questions left.

1) Am I right to suppose that the standard deviation you are refering to is the choletsky decomposition standard deviation, which should be the first number in the unscaled IRFs? If not, where can I find the right standard deviation?

2) As far as I understand the add in, the scaling option as well as the variance decomposition and the save IRFS to matrix option are only available in the drop down menu and not via a code. At least in the guide they are not mentioned. Could you provide the possible comands, such that I dont have to use the drop down menu every time?

Regarding the Unit Root:

Thanks so much for clearing that up. If Sims did that, I should also be allowed to do that!

Best,
Ben

[dakila]

1) the standard deviation I am referring is the standard deviation of FFR, not cholesky decomposition of VAR residual VCMatrix.
For instance, you can estimate it by command of @stdevp(ffr) before variable transformation.

2) favar(factor=3,horizon=48,rep=1000,ci=0.9,save=1,vd=1, scale=1, sn=0.07816)

save =1 (0) save the IRF.
vd =1 (0) - variance decomposition analysis.
scale = 1 (0) - scale the IRF.
sn - scale factor.

[Bens]

You are the best! Thanks for your help and have a good day

[Bens]

Hey dakila,

am I right to suppose that your great add-in has this feature that it uses accumulated irfs for variables that are log-differentiated and otherwhise it uses the standard irfs? Regarding the interpretation. The units on the axes are standard deviation units or % units? Sorry for asking such a dump question, i just wanna make sure I get your add-in right

[dakila]

Hi Ben,

am I right to suppose that your great add-in has this feature that it uses accumulated irfs for variables that are log-differentiated and otherwhise it uses the standard irfs?

Yes, you are right.

The units on the axes are standard deviation units or % units?

standard deviation units because the all variables are standardized.

[Bens]

Thanks so much for your response. Standard deviation units if all variables are standardized means that a 1 standard deviation increase is also a 1 unit increase? So if the Impulse of a log differentiated variables shows a 0.5 unit increase, does that mean a 0.5 percent increase? Sorry for asking such questions...first time working with a favar

[Bens]

ps. Which horizon did you choose for the forecast error variance decomposition? Bernanke et al. chose 60 month. Did you also do that ?

[dakila]

No. All interpretations are in std. its does not matter whether the horizon is 48 or 60. The result is the same.

[Bens]

Hey dakila,

thanks for your help. I am sorry that I have to ask all those questions, but since I am not able to look into your code (which I totally understand) I have to ask manualy. I am currently performing residual diagnostics on the residuals of the estimated FAVARs called favarb01 and favar01. One is used to get accumulated IRFS while the other is used for standard IRFS depending on the variables. I am currious why the favarb01 takes different inputs than the favar01. While the favar01 takes the standard rotated factors (Facrot_1 etc) the other uses some transformated version of those factors (Facrot_1_0 etc.). Those factors look similar in the start but then continue to differ. What is the reason for that? What do you think: Which one of the estimaed FAVAR should I use for residual diagnostics?

Best,
Ben

[dakila]

Sorry. I don't understand your question. How did you get favarb01 object?

[Bens]

Sorry if I didnt explain it properly. I use eviews 9 and ran your add-in ( I think I havent had this problem when using Eviews 10 in University). I simply updated the dataset and did not change anything. I have two var models in the final output. One is called favar01, one is called favarb01. If I click on "impulse" favar01 has the "accumulated impulse responses" box activated, while favarb01 doesnt. Favar01 also uses inputs called _factrot1 etc. while the favarb01 uses inputs called factor1_0. I am wondering about the difference between factor1_0 and factor1 and which of the favar models (favar01 and favarb01) is in the end used for the IRFS. Furthermore: When using Eviews9 I allways get a syntax error at the end of the code, but it seems like its not harming the calculations.

Best,
Ben