The RATS Software Forum

Hello
i am working on a garch model with detailed instructions, all goes well but when i execute the last maximasation procedure, it can't find the data points.

## SR10. Missing Values And/Or SMPL Options Leave No Usable Data Points.

Have you read page UG-294 in the version 8 User's Guide?

TomDoan wrote:Have you read page UG-294 in the version 8 User's Guide?

dear Tom

i've seen the users' guide after you told me.i've rechecked all data points and i've found 4 NA in the last day, and i have tried with a short data period without this day but it still doesn't work. Is there any possiblility that i could delete this last data which is generated by the procedure before?

Is there a reason you're not using the GARCH instruction?

The problem is that you're using a bad set of guess values. Your off-diagonal GARCH terms are more persistent than your diagonal GARCH terms---as a result, you're producing a non-positive definite covariance matrix on the very first entry.

TomDoan wrote:Is there a reason you're not using the GARCH instruction?

The problem is that you're using a bad set of guess values. Your off-diagonal GARCH terms are more persistent than your diagonal GARCH terms---as a result, you're producing a non-positive definite covariance matrix on the very first entry.

dear Tom
thank you very much. i am not using GARCH instruction because there is a paper written by BROOKS who compared different results of a GARCH model by different software, RATS, EVIEWS, GAUSS. and there is a quite difference. So i cant really say who is better. However, with these detailed instructions, i know what i do. Anyway, gratitude Tom. you're great.

dear Tom

i don't know whether you still remerber my problem, it has been successfully solved by you. now the new problem is when i put two lines of new code in the old code who worked well, unfortunately it doesnt work any more. here is the modified part of the code:

frml H11 = c11 + b11 * U11{1} + b12 * U22{1} + b13 * U33{1} + a11 * r1{1}**2
frml H22 = c22 + b21 * U11{1} + b22 * U22{1} + b23 * U33{1} + a22 * r2{1}**2
frml H33 = c33 + b31 * U11{1} + b32 * U22{1} + b33 * U33{1} + a33 * r3{1}**2


and the old code was this:

frml H11 = c11 + b11 * U11{1} + a11 * r1{1}**2
frml H22 = c22 + + b22 * U22{1} + a22 * r2{1}**2
frml H33 = c33 + + b33 * U33{1} + a33 * r3{1}**2

i've upgrade the new code in the attachement, and with this new code, it can't CONVERGE.and i have try really many times with different guess values

You still are using very bad guess values. You have .05 coefficients on the own lagged variance terms and .50 on the lagged squared residuals. In a typical GARCH, the lagged variance term is the dominant one. Plus, the guess values on the off-diagonals give an unstable recursion.

Have you looked at the behavior of univariate GARCH models on these series first? Given that what you're doing nests uni GARCH models, starting with guess values based upon the univariates would be a good idea.

TomDoan wrote:You still are using very bad guess values. You have .05 coefficients on the own lagged variance terms and .50 on the lagged squared residuals. In a typical GARCH, the lagged variance term is the dominant one. Plus, the guess values on the off-diagonals give an unstable recursion.

Have you looked at the behavior of univariate GARCH models on these series first? Given that what you're doing nests uni GARCH models, starting with guess values based upon the univariates would be a good idea.

dear Tom,

i did what you've told me to do. i've tried with univarite GARCH .the coefficients as you said, are all aroud 0.9, so i've changed the lagged variance to 0.9 and 0.05 for the lagged squared residuals. for the co-variance, i set 0.6. unfortunately, still, that doesnt work. is there any possibility that is not the problem , dear Tom?

You are restricting your estimation range to a three year period when a DVECH-type model really doesn't work well. Whether any relatively simple MV GARCH model can properly handle those data is unclear, but a DVECH, with the separately estimated covariance equations, definitely can't.