The RATS Software Forum

No problem.

I also tried to roughly recreate the results of a paper that uses the 4th technique, the same thing happens as when I use may own data series. Here I did not change any of the code and the data is fairly short time series.

I have enclosed the data and also the code; although there is basically no change in it.

Thanks.

The original data set had gstart=2 because the data had to be converted to returns. With your data set, you get gstart=1 since you already have return data which doesn't work on the AGDCC model because there is no room for the lags. Instead, use an entry range of 2 to gend rather than gstart to gend on the MAXIMIZE.

Sorry, stupid, I missed that with this test.

Turns out the problem with the other data was also to do with my wrong returns convertion in the code.

Thanks a lot, I think I was over thinking things.

Hey, thanks for this code. I have been just been using it to produce some results for some ADCC and AGDCC models. I was wondering if anyone could help me understand something.

When I run a 3/ 4 time series model I can understand all the results for all four techniques but when I run it for only 2 series I get 6 coefficients for example alpha 1 and alpha 2 using AGDCC. Looking at the original article I can understand ADCC as a model. However; I am not sure the generalization’s impact and why that leads to the two coefficients for each parameter being produced. How do I interpret these? Or alternatively is the point that I should not be using the generalized techniques when looking at the relationship between only two series?

Thank you.

HenleyPhD wrote:Hey, thanks for this code. I have been just been using it to produce some results for some ADCC and AGDCC models. I was wondering if anyone could help me understand something.

When I run a 3/ 4 time series model I can understand all the results for all four techniques but when I run it for only 2 series I get 6 coefficients for example alpha 1 and alpha 2 using AGDCC. Looking at the original article I can understand ADCC as a model. However; I am not sure the generalization’s impact and why that leads to the two coefficients for each parameter being produced. How do I interpret these? Or alternatively is the point that I should not be using the generalized techniques when looking at the relationship between only two series?

Thank you.

It looks like it needs a little help moving off the zero guess values for the GQ matrices. Add the PMETHOD=SIMPLEX,PITERS=5 options to the last two MAXIMIZE instructions:

Code: Select all

maximize(start=%(StartQC()),pmethod=simplex,piters=5,method=bfgs,title="GDDCC") logl gstart gend
disp "GDDCC BIC" -2.0*%logl+(%nreg+uniparms)*log(%nobs)
*
* AGDDCC (Asymmetric Generalized Diagonal DCC)
*
compute gq=%const(0.0)
nonlin aq bq gq
maximize(start=%(StartQC()),pmethod=simplex,piters=5,method=bfgs,title="AGDDCC") logl gstart gend
disp "AGDDCC BIC" -2.0*%logl+(%nreg+uniparms)*log(%nobs)

Tom,

Thank you, those options for the MAXIMIZE instruction did help for a lot of my models.

Could you possibility help me understand two others things? If I try and run the code for shorter data series (300/ 400 obvs) I find that the results are very sensitive to the initial values specified in the code. This does make sense to me; would it then be correct to base the starting values on the univariate GARCH results? In this particular case these tend to lower than with most of the other models that I have run.

Also in one of the models I have tried I end up with one negative alpha value, is that wrong? Should I try and re-run it using alternative options?

HenleyPhD wrote:Tom,

Thank you, those options for the MAXIMIZE instruction did help for a lot of my models.

Could you possibility help me understand two others things? If I try and run the code for shorter data series (300/ 400 obvs) I find that the results are very sensitive to the initial values specified in the code. This does make sense to me; would it then be correct to base the starting values on the univariate GARCH results? In this particular case these tend to lower than with most of the other models that I have run.

Also in one of the models I have tried I end up with one negative alpha value, is that wrong? Should I try and re-run it using alternative options?

It's quite possible that 300/400 data points aren't enough to fit a model like that reliably.

There's nothing wrong with a negative alpha since the alpha's enter in "squares". In fact, if you flip the sign of one of the input series, you would get the same likelihood, but with several coefficients changing signs, including the alpha on the affected series.