How to get the ideal result in a FIEC-HYGARCH Model?

[u8810137]

Hi all,

I've tried many times to deal with FIEC-HYGARCH Model (4 variables, lags=5), but none of the results meet my requirement. This model is used for analyzing the spillover effect and long memory.

Sometimes I get the message: "NO CONVERGENCE IN XX ITERATIONS
LAST CRITERION WAS 0.0000000
ESTIMATION POSSIBLY HAS STALLED OR MACHINE ROUNDOFF IS MAKING FURTHER PROGRESS DIFFICULT
TRY HIGHER SUBITERATIONS LIMIT, TIGHTER CVCRIT, DIFFERENT SETTING FOR EXACTLINE OR ALPHA ON NLPAR
RESTARTING ESTIMATION FROM LAST ESTIMATES OR DIFFERENT INITIAL GUESSES MIGHT ALSO WORK"

Even I get the result of convergence under the BHHH method, the d (fractional differencing coefficient) is not reasonable, and many coefficients have high p-value, like 0.8 or 0.9.

Perhaps the reason is that there are too much parameters in this model, but it's necessary by my professor's requirement.
And I have no idea in setting the number of "iters", "subiters", are there any rules between them, such as percentage?
Or do I need to adjust the number after LOGL under the maximize instruction?
Or is there anything wrong with my initial guess under the simplex method?
(My initial guess under the simplex method is f1~f12:0.1; df11~df34:0.3; g11~g34:0.5.)

So, can any of you help me to figure this problem out?
Thanks in advance!

[TomDoan]

Hi all,

I've tried many times to deal with FIEC-HYGARCH Model (4 variables, lags=5), but none of the results meet my requirement. This model is used for analyzing the spillover effect and long memory.

Sometimes I get the message: "NO CONVERGENCE IN XX ITERATIONS
LAST CRITERION WAS 0.0000000
ESTIMATION POSSIBLY HAS STALLED OR MACHINE ROUNDOFF IS MAKING FURTHER PROGRESS DIFFICULT
TRY HIGHER SUBITERATIONS LIMIT, TIGHTER CVCRIT, DIFFERENT SETTING FOR EXACTLINE OR ALPHA ON NLPAR
RESTARTING ESTIMATION FROM LAST ESTIMATES OR DIFFERENT INITIAL GUESSES MIGHT ALSO WORK"

Even I get the result of convergence under the BHHH method, the d (fractional differencing coefficient) is not reasonable, and many coefficients have high p-value, like 0.8 or 0.9.

Perhaps the reason is that there are too much parameters in this model, but it's necessary by my professor's requirement.
And I have no idea in setting the number of "iters", "subiters", are there any rules between them, such as percentage?
Or do I need to adjust the number after LOGL under the maximize instruction?
Or is there anything wrong with my initial guess under the simplex method?
(My initial guess under the simplex method is f1~f12:0.1; df11~df34:0.3; g11~g34:0.5.)

So, can any of you help me to figure this problem out?
Thanks in advance!

First, you can improve the general behavior of the estimation process by scaling up the log series by a factor of 100. That changes the scale of the VC parameters by 10000, which makes them more on the order of 1, rather than 1.e-4.

set lntes = 100.0*log(tes)
set lntef = 100.0*log(tef)
set lntfs = 100.0*log(tfs)
set lntff = 100.0*log(tff)

Second, you can start with the mean parameters fixed and just estimate the GARCH parameters. Once you get that to converge you can put the mean parameters back in.

Third, you are doing the fractional differencing with the default 0 padding. When you're doing that on a strictly positive series (such as the uu(i,i)), that's not a good choice for the padding. Using a "presample" equal to the unconditional variance or covariance of the residuals would be better. (See, for instance, the Baillie-Bollerslev FIGARCH replication).

One thing to note, however, is that you have a very complicated GARCH model for the variance with nothing such as BEKK or CC restrictions forcing it to be positive definite. It can be very difficult to estimate parameters when you have boundary issues like this model will have.

[u8810137]

Hi all,

I've tried many times to deal with FIEC-HYGARCH Model (4 variables, lags=5), but none of the results meet my requirement. This model is used for analyzing the spillover effect and long memory.

Sometimes I get the message: "NO CONVERGENCE IN XX ITERATIONS
LAST CRITERION WAS 0.0000000
ESTIMATION POSSIBLY HAS STALLED OR MACHINE ROUNDOFF IS MAKING FURTHER PROGRESS DIFFICULT
TRY HIGHER SUBITERATIONS LIMIT, TIGHTER CVCRIT, DIFFERENT SETTING FOR EXACTLINE OR ALPHA ON NLPAR
RESTARTING ESTIMATION FROM LAST ESTIMATES OR DIFFERENT INITIAL GUESSES MIGHT ALSO WORK"

Even I get the result of convergence under the BHHH method, the d (fractional differencing coefficient) is not reasonable, and many coefficients have high p-value, like 0.8 or 0.9.

Perhaps the reason is that there are too much parameters in this model, but it's necessary by my professor's requirement.
And I have no idea in setting the number of "iters", "subiters", are there any rules between them, such as percentage?
Or do I need to adjust the number after LOGL under the maximize instruction?
Or is there anything wrong with my initial guess under the simplex method?
(My initial guess under the simplex method is f1~f12:0.1; df11~df34:0.3; g11~g34:0.5.)

So, can any of you help me to figure this problem out?
Thanks in advance!

First, you can improve the general behavior of the estimation process by scaling up the log series by a factor of 100. That changes the scale of the VC parameters by 10000, which makes them more on the order of 1, rather than 1.e-4.

set lntes = 100.0*log(tes)
set lntef = 100.0*log(tef)
set lntfs = 100.0*log(tfs)
set lntff = 100.0*log(tff)

Second, you can start with the mean parameters fixed and just estimate the GARCH parameters. Once you get that to converge you can put the mean parameters back in.

Third, you are doing the fractional differencing with the default 0 padding. When you're doing that on a strictly positive series (such as the uu(i,i)), that's not a good choice for the padding. Using a "presample" equal to the unconditional variance or covariance of the residuals would be better. (See, for instance, the Baillie-Bollerslev FIGARCH replication).

One thing to note, however, is that you have a very complicated GARCH model for the variance with nothing such as BEKK or CC restrictions forcing it to be positive definite. It can be very difficult to estimate parameters when you have boundary issues like this model will have.

I've tried the first and third methods, but they didn't work.
And I don't quite know how to do your second method in rats code.
Can you give me more details?

Do you think it's a good way to try different initial guess, and get different simplex output; then, put the output to the bhhh method?

Thanks!

[TomDoan]

Do MAXIMIZE with PARMSET=GARCHPARMS rather than PARMSET=MEANPARMS+GARCHPARMS. Once that works, go back to the combination parmset.

I would always do preliminary simplex iterations for a model like this. However, instead of running the simplex, saving the estimates and re-entering them, it's much simpler to use generic guess values and use the PMETHOD=SIMPLEX option along with METHOD=BFGS. I wouldn't recommend using BHHH except as a final step to get BHHH standard errors.

Have you read section 4.15 ("Troubleshooting") of the RATS User's Guide? That's why it's there.

[u8810137]

What do you mean by "generic guess values"?
My initial guess were set by myself, is it wrong?

I've read that part(trouble shooting), but I still don't get it!

Or is there anything wrong with my rats code for this model?

Thanks a lot!

[TomDoan]

I mean guess values like .8, .2 ... that you think are in the right range.

There's nothing wrong with your program. For many reasons, that's a very hard model to fit. Have you tried a simpler model first? (Fewer lags, common d).

[u8810137]

So I don't need to change the initial values all the time in order to get good results under the bhhh/bfgs method, right?

sometimes it converges, but some coefficients have zero standard error, t value and p-value.

[u8810137]

sorry, one more question
Are the coefficients of simplex output as same as that of bhhh/bfgs output?
In my case, both of them are same.
I think that's weird.

thanks.

[TomDoan]

sorry, one more question
Are the coefficients of simplex output as same as that of bhhh/bfgs output?
In my case, both of them are same.
I think that's weird.

thanks.

If you do PMETHOD=SIMPLEX, there isn't any output produced by the simplex iterations---it just internally switches to the other algorithm once the simplex iterations are done. If you do any method to convergence and then re-do with the MAXIMIZE with a different method, the coefficients won't change (noticeably, if at all).

[TomDoan]

So I don't need to change the initial values all the time in order to get good results under the bhhh/bfgs method, right?

sometimes it converges, but some coefficients have zero standard error, t value and p-value.

The preliminary simplex iterations aren't as sensitive to the choice of guess values, but you still need to be on the correct "hill" if there are multiple modes.

[u8810137]

Hi, I'm back wih a new question.
I don't quite understand how to add dummy variable into the model.
Firstly, I need to add one dummy variable (tp) into the mean equations, and I've already set i or 0 in the excel file.
But when I run it (attached file), it shows "## SR10. Missing Values And/Or SMPL Options Leave No Usable Data Points"
Can you point out my problem?

Thanks again!

[TomDoan]

You attached the wrong data file. The program uses data9.xls.

[TomDoan]

You can simplify your program considerably using the ideas from

http://www.estima.com/forum/viewtopic.php?f=11&t=1460

Your mean model is quite large, but can be specified and handled flexibly using a "VAR" system with the Z as an added "DETERMINISTIC" variable.

[u8810137]

Sorry for my mistake! Actually data8 is as same as data9.
Thank you for your quick reply and advice!
I'll try to make it out!

Regards