BVARs yet again

[donihue]

I apologise for once again returning to this subject!

The GUI works perfectly, but I am having problems using the command-line interface for this excellent Add-in (I want to use a loop to calculate a range of models over different sets of hyperparameters).

According to the .pdf file for this Add-in, the command line should read (as an example) "bvar(Prior=0) gdp interest inflation", with other options available, such as L0, L1, etc

I have tried

bvar(Prior=0) mb gdp usd_rate cpi
bvar(Prior=0,L0=0.8,L1=0.1,L3=2,L4=1,L5=0.07,Mu5=0.1,Mu6=0.1,QM=12) mb gdp usd_rate cpi
bvar(Prior=0 L0=0.8 L1=0.1 L3=2 L4=1 L5=0.07 Mu5=0.1 Mu6=0.1 QM=12) mb gdp usd_rate cpi
bvar(Prior=0,lambda0=0.8,lambda1=0.1,lambda3=2,lambda4=1,mu5=0.1,mu6=0.1,qm=12) mb gdp usd_rate cpi

but in all cases, received the cryptic message "Syntax error".

This clearly indicates that I am doing something obviously wrong in formulating the command, but I cannot see what. Can you help, please?

Regards
Donihue

[EViews Esther]

It was my bad. The first argument should be the number of lags.

For example, you can say:
bvar(Prior=0) 2 mb gdp usd_rate cpi
(Note: "bvar(Prior=0) mb gdp usd_rate cpi" will report the error message ).

When you re-install the BVAR addin, you will see the updated command options.

[donihue]

Many thanks, Esther - it works fine now

regards
Donihue

[donihue]

Hello again, Esther,

Any chance of adding a command-line option to produce a model, as in the GUI?
Given the lack of a direct forecast command for BVARs (as also for VECMs) this is essential for forecasting.

Again, many thanks for your work on this Add-in: it is a real life-saver for we BVAR workers!

Regards
Donihue

[mayxanh]

Hi Esther,

I have just installed BVAR and use the exact command that you shown above. However, there is an error message keeping poping up: "_SH01 is not defined" Could you please to explain me why? Thank you very much.

T

[EViews Esther]

It seems that there exist a singularity problem. Could you upload your workfile? Then, I will figure out the solution.

[mayxanh]

Esther,

Thank you for your assistance. I upload the workfile named test.

When I run the following command:

bvar(Prior=0) 2 LOG_IRR_SA_CA LOG_CCLI_S IR_SPREAD_U LOG_GDP_CHN_SA LOG_CPI_TOTAL_S LOG_CERI_S

The error message is "_SH01 is not defined"

When I run the command:

bvar(Prior=0,m=model) 2 LOG_IRR_SA_CA D(LOG_CCLI_S) D(IR_SPREAD_U) D(LOG_GDP_CHN_SA) D(LOG_CPI_TOTAL_S) D(LOG_CERI_S)

(because all series are non-stationary, except LOG_IRR_SA_CA)

then the model is estimated without any error message. However, there are a series of error messages inside the model object:

"syntax error in ..."

I need to get the model object in order to do the forecasting using rolling windows.

Thank you again

[EViews Esther]

Let me clarify the problem first. Since the given regressors are perfectly collinear, the prior values for the residual covariance cannot be calculated.

Following the paper (Sims and Zha; 1998), the current addin calculates the diagonal elements of the prior for the residual covariance matrix S (see Equation 10 in the documentation) by OLS steps.
For example, the 1st diagonal element of S for the 1st lag of LOG_CCLI_S is derived by

Code: Select all

!m = 3 'number of variables
matrix(!m,!m) S 'storage for S
equation EQ1.ls D(LOG_CCLI_S) D(LOG_CCLI_S)(-1) c  'Do OLS: LOG_CCLI_S on D(LOG_CCLI_S)(-1) and c only.
!SSR = EQ01.@ssr
s(1,1) = !SSR/EQ1.@regobs

When you have perfect collinearity problem in OLS steps, the prior for the residual covariance cannot be calculated.

I think that the solution to your problem can be resolved by changing the prior calculation steps. For example,

Code: Select all

equation EQ1.ls LOG_IRR_SA_CA D(LOG_CCLI_S) D(IR_SPREAD_U) D(LOG_GDP_CHN_SA) D(LOG_CPI_TOTAL_S) D(LOG_CERI_S) c
                                                               'Do OLS: all regressors are used
matrix S = EQ1.@coefcov

[mayxanh]

Hi Esther,

Thank you for your explanation. Regarding the problem with the makemodel, I find out that the problem arise from the usage of D(x)(-1) as the variable so that the @word function does not return the correct word. It first returns D(x) then (-1) instead of returning the whole string D(x)(-1).

I wrote a fix for the subroutine eqmake and this time it works perfectly.


subroutine local eqmake(string %spec, string %coefs, string %speceq)
%speceq = ""

!start = 1
!length = @LENGTH(%spec)
!end = 1

for !i=1 to @wcount(%coefs)
'%word = @word(%spec,!i)

!end = @INSTR (%spec, " ", !i)

if !end =0 or !end = !length then
!end = !length + 1
endif

!size = !end - !start

%word = @MID(%spec, !start, !size)

if @upper(%word) = "C" then 'constant
%word = ""
endif
%coef = @word(%coefs,!i)
if @val(@word(%coefs,!i))>=0 then
%speceq = %speceq+"+"+%coef+"*"+%word
else
%coef=@right(%coef,@len(%coef)-1)
%speceq = %speceq+"-"+%coef+"*"+%word
endif

!start = !end + 1

next


if @val(@word(%coefs,1))>=0 then
%speceq = @right(%speceq,@len(%speceq)-1)
endif
%speceq = @left(%speceq,@len(%speceq)-1)
endsub

[mayxanh]

Hi Esther,

Another question I would like to ask is how to select the lag length for a BVAR? In a VAR system, we can use Information criterion to automatically select the lag length. How's about BVAR?


Thank you

[EViews Gareth]

That's a little outside of the realm of technical support.

[mayxanh]

Hi Gareth, Esther,

Thanks a lot for your help. Another question I would like to ask is that you use OLS to estimate the BVAR? The VAR system is also estimated by OLS? Thanks again.

[EViews Gareth]

No. The BVAR uses Bayesian techniques, as outlined in the documentation.

[mjfl]

hi,

im new to BVAR and would like to ask if there is any command to forecast the bvar?
i see that making the model helps but can someone pls enlighten how to forecast the bvar directly or from the model?
i cant understand what is
1) no of forecast period
2) forecast selection

becasue even after trying with or w/o it my workfile remains same without any forecast. pls help as im really trying to learn and though the query sound really silly.

[EViews Gareth]

Once you have the model, just solve the model over the period you wish to forecast.