Codes and output for a VECT model

[anozman]

Hi Tom,

I checked the ECT.RPF example file's output, I was wondering whether you could help me to understand two issues:

1. Why the dependent variable for one of the VAR equations is FTBS3, not the D_FTBS3?
2. I tried to use one lag in the Cointegration relation, and set up a one-lag VECT. unless i set the lags in the system as below, I could only get two ECT{1} terms. could you please tell me why?

system(model=ect2model)
variables ftbs3 ftb12 fcm7
lags 1 to 2 <-------------------
ect ect1 ect2
end(system)
estimate
errors(model=ect2model,steps=36)



This is the RATS output for question 1

VAR/System - Estimation by Cointegrated Least Squares
Monthly Data From 1975:07 To 2001:06
Usable Observations 312

Dependent Variable FTBS3
Mean of Dependent Variable -0.005929487
Std Error of Dependent Variable 0.572477918
Standard Error of Estimate 0.491510475
Sum of Squared Residuals 71.508433993
Durbin-Watson Statistic 1.9616

Variable Coeff Std Error T-Stat Signif
************************************************************************************
1. D_FTBS3{1} 0.143892238 0.183591916 0.78376 0.43380689
2. D_FTBS3{2} 0.192186442 0.183710031 1.04614 0.29634985
3. D_FTBS3{3} -0.092909875 0.191610137 -0.48489 0.62811289
4. D_FTBS3{4} 0.408247277 0.174242798 2.34298 0.01979251
5. D_FTBS3{5} -0.212660827 0.169932295 -1.25144 0.21176048
6. D_FTB12{1} 0.290088325 0.260050689 1.11551 0.26553809
7. D_FTB12{2} -0.362872849 0.253978857 -1.42875 0.15412986
8. D_FTB12{3} 0.063761999 0.259907946 0.24533 0.80637455
9. D_FTB12{4} -0.577081483 0.246168719 -2.34425 0.01972637
10. D_FTB12{5} 0.151509694 0.247373031 0.61247 0.54069404
11. D_FCM7{1} 0.218441482 0.191699666 1.13950 0.25541657
12. D_FCM7{2} -0.278771615 0.195277984 -1.42756 0.15447181
13. D_FCM7{3} 0.165902869 0.196382770 0.84479 0.39890849
14. D_FCM7{4} 0.162678086 0.199755989 0.81438 0.41607968
15. D_FCM7{5} 0.233716355 0.196062364 1.19205 0.23419602
16. EC1{1} 0.062282194 0.027826301 2.23825 0.02594825

Regards,
anozman

[TomDoan]

That's described in "Setting up the VAR" in Section 7.8 of the User's Guide. You set up the VAR using the undifferenced variables and the number of lags in the VAR on the original lags, which will be one more than the number of lags on the differences. ESTIMATE takes care of converting that into the restricted form.

[anozman]

Thank you very much!

[anozman]

Hi Tom,

Do you know how to normalize and show the cointegration relation equation specification using the above program? Or do I have to use CATS2 to do that?

Many thanks,
Anozman

[TomDoan]

If you just need a normalization, then you can do that with RATS. (If you want to test overidentifying restrictions, you would need CATS). For an example of a bivariate normalization, see the MHHP682.RPF example from the Martin-Hurn-Harris textbook examples.

[anozman]

Hi everyone,

If I use the @johmle procedure and find three cointegration relations. Hoe can I find the specifications of these three equations. It seems the default setting for @johmle is to only show the specification for the largest eigenvalue cointegration vector.

Many thanks,
Anozman

[TomDoan]

The VECTORS option gives you the full set of cointegrating vectors. There's a revised version of ECT.RPF with two cointegrating vectors estimated by @JOHMLE in the RATS v9 User's Guide.

[anozman]

Thank you very much for the information.

How do I normalise the three coefficients in three cointegration vectors using @johmle? Does it really matter to normalise the coefficients in cointegration vecotrs and what coefficient normalisation impacts on the variance decomposition results for the final VECT model?

Many thanks,
Anozman

[TomDoan]

Thank you very much for the information.

How do I normalise the three coefficients in three cointegration vectors using @johmle?

That was answered above.

Does it really matter to normalise the coefficients in cointegration vecotrs and what coefficient normalisation impacts on the variance decomposition results for the final VECT model?

No, you don't. Normalizing the cointegrating vectors is only required if you're trying to interpret them and their loadings. It has no effect on any forecast-type calculation like impulse responses or FEVD.

[anozman]

Thank you very much for this feedback. I tried to use @varlagselect with AIC, SBC, BIC and GTOS to select lag length. For the first 3 criteria, I got lag 0 as the best selection and for GTOS, I got lag 4 as the chosen one. Could any one tell me how I could get lag 0 as the best option from 3 information criteria and which lag length I should choose given these contradicting outcomes?

Regards,
Anozman

[TomDoan]

Thank you very much for this feedback. I tried to use @varlagselect with AIC, SBC, BIC and GTOS to select lag length. For the first 3 criteria, I got lag 0 as the best selection and for GTOS, I got lag 4 as the chosen one. Could any one tell me how I could get lag 0 as the best option from 3 information criteria and which lag length I should choose given these contradicting outcomes?

Is this a VAR on the data on which you're trying to run an ECT? Are you using @VARLAGSELECT on the levels or differences? If it's the levels, then it seems hard to believe that you got a unit root test to accept unit roots. If it's on the differences, then your VAR is misspecified if the variables are cointegrated.

If the process is vector white noise, or something close to it, the optimal lags should be zero. If, for instance, you take a VAR from a set of stock returns, you would probably expect the lag length criteria to pick zero.

[anozman]

The series are first differenced and log transformed time series. In addition, there isn't any theoretical justification for any real cointegration relationships. If data shows some cointegrations, then it is most likely to be spurious. I was just wondering when 3 out of 4 criteria suggest 0 lag and one (GTOS) suggests 4 lags, what lag length should I use? If I use lag 0, how to set up the VAR. I was thinking about using just one lag for a smallest VAR.

Regards,
Anozman

[TomDoan]

GTOS isn't a good choice for a VAR. If the AIC is coming in zero as well, it sounds like zero is the number. If your goal is to analyze the levels rather than the differences, the best way to handle it is to put the variables in levels, LAGS at 1, and put an empty ECT instruction---that will produce the same thing as a "VAR" with no lags on the differences, but will analyze out as levels if you do forecasts or impulse responses. If you really want to do the VAR in differences, just leave the LAGS instruction out.

[anozman]

Thank you very much for your feedback. This VAR setting will be used to produce the spillover tables and the spillover index based on Diebold and Yilmaz 2009. If I use lag 0, Does it make any conceptual sense ?

[TomDoan]

Thank you very much for your feedback. This VAR setting will be used to produce the spillover tables and the spillover index based on Diebold and Yilmaz 2009. If I use lag 0, Does it make any conceptual sense ?

Diebold-Yilmaz is a calculation that focuses on the effects of changing the contemporaneous model. The usefulness of that doesn't depend upon the dynamics. All that changes is that there is no difference between the spillover measure that depends upon the forecast horizon, since those will all be the same.