Bivariate VAR(4) GARCH(1,1)-BEKK model quality

[elisabettac]

Hi, I am trying to estimate an asymmetric bivariate GARCH-BEKK model with as mean equation a VAR(4).
I run the code posted below on with the dataset attached (they are log returns*100 of price series of stock indices), but the majority of the coefficients estimated are not significant. Can somebody help me understand how to improve the quality of such estimations?

Thanks

Code:
OPEN DATA "/Users/E/Desktop/Analysis/Bivariate GARCH/US.xlsx"
CALENDAR(D) 1991:1:1
DATA(FORMAT=XLSX,ORG=COLUMNS) 1991:01:01 2016:12:30 CHINA US
*
SYSTEM(MODEL=VAR1)
VARIABLES CHINA US
LAGS 1 TO 4
DET Constant
END(SYSTEM)
*
GARCH(P=1,Q=1,MODEL=VAR1,MV=BEKK,ASYMMETRIC,ROBUSTERRORS,PMETHOD=SIMPLEX,PITERS=20,METHOD=BFGS,ITERS=50) / CHINA US

[TomDoan]

A lot of the mean model coefficients are insignificant, but that's not unexpected if you do a 4 lag VAR. With 6000+ data points, it's not like the model is going to suffer much from having 18 mean parameters. If you're talking about the GARCH coefficients, if the two variance processes don't have much to do with each other (and apparently they don't), there's nothing you can do---sometimes zero is the correct answer.

[elisabettac]

A lot of the mean model coefficients are insignificant, but that's not unexpected if you do a 4 lag VAR. With 6000+ data points, it's not like the model is going to suffer much from having 18 mean parameters. If you're talking about the GARCH coefficients, if the two variance processes don't have much to do with each other (and apparently they don't), there's nothing you can do---sometimes zero is the correct answer.

Thanks a lot for your help, Tom.
So would you recommend changing the number of lags in the mean equation? If so, would a higher number be better?

[TomDoan]

No, probably fewer if anything. It's usually best to start small with the mean model and add lags only if the serial correlation diagnostics come back significant.

[elisabettac]

No, probably fewer if anything. It's usually best to start small with the mean model and add lags only if the serial correlation diagnostics come back significant.

Okay perfect, thanks a lot.
One final question, among the last part of the code for the GARCH model ("ROBUSTERRORS,PMETHOD=SIMPLEX,PITERS=20,METHOD=BFGS,ITERS=50"), is there anything you would add/remove/improve?

[TomDoan]

Is that a typo? Do you mean ITERS=50 or ITERS=500? If the former, there is no reason to cut down on the number of main iterations---if it converges quicker than the default 200, that's great. The only reason we put a limit on ITERS is so it won't run for hours getting nowhere if the model has problems. PITERS is a different situation, since you don't want to overdo those---in general, it's a waste of computing time, since the preliminary methods are quite a bit less efficient than BFGS once the parameters are in the correct part of the likelihood surface. 20 is OK---I generally try 10 to start, but if you've been using 20 and getting convergence, I wouldn't change that.

[elisabettac]

Is that a typo? Do you mean ITERS=50 or ITERS=500? If the former, there is no reason to cut down on the number of main iterations---if it converges quicker than the default 200, that's great. The only reason we put a limit on ITERS is so it won't run for hours getting nowhere if the model has problems. PITERS is a different situation, since you don't want to overdo those---in general, it's a waste of computing time, since the preliminary methods are quite a bit less efficient than BFGS once the parameters are in the correct part of the likelihood surface. 20 is OK---I generally try 10 to start, but if you've been using 20 and getting convergence, I wouldn't change that.

Okay perfect. Thanks a lot for your help, Tom!