Chan & Maheu, JBES 2002 (Jump GARCH model)

[Max Resende]

Dear RATS users,

I have been working with Chan & Maheu, JBES 2002 (Jump GARCH model) for a while and I have bumped into some issues:

1) I have 2008 observations; estimations are only using 1767...why does that happen?

2) my estimation routine has failed to converge upon an optimum so the parameter estimates and their standard errors may be unreliable. How can i fics that? I have changed the parameters value and everything that I could think of, but I was unable to fics the problem.

3) Does the converge problem has something to do with the estimations (that seems to converge) not using all the available data?

I have attached both my data file and code.

I hope you guys can help me.

Max

[TomDoan]

This will not get you the "end" of 2017---it's the 12th entry within 2017 (which is 1770)

end2017=2017:12

To get the last date in 2017, you want (2018:1)-1. Note, however, that the Chan and Maheu paper does subranges because that's appropriate for their dataset and for what they are trying to show. If you're just trying to fit a jump GARCH model, you don't want to monkey with year-based subranges.

[Max Resende]

Dear Tom Dam,

It works just fine. Thank you once again.

But now I have been trying to forecast the last 30 periods of data of for the last model ARJI-h(t), but I couldn't do it. This is the first time that I try to do out of sample forecasts. How can I get the MSE for the forecast?

I have attached my data and the code that I have been using so far.

Thank you once again.

Max

[TomDoan]

I don't believe we've ever provided anything to do multi-step forecasts (one-step is relatively easy) and I don't know if a closed-form expression even exists for multi-step forecasts.

[Max Resende]

Tom Dam,

What I had in mind would be something like we have on the RATS Handbook for ARCH/GARCH and Volatility Models pgs 57-58 and applied the @GACRGFORE command.

Please, correct me if I am wrong: Because I have a daily data set, it doesn't make sense to forecast a 1-day forecast (E(R)=0). It would make sense if I had a monthly or quarterly data set.

Once again, thank you!!

[TomDoan]

Closed form forecasts (beyond one-step) exist for only certain types of GARCH models---ones where the law of iterated expectations apply to the out-of-sample residuals terms. (One-step generally isn't a problem since the lagged residuals are treated as "observable").

[prashantj]

Hi Tom,
I tried to use jumpgarch.src file without modification, but it didn’t estimate the model. It mentions in the help that There should be no need to edit this if you intend to use the models described in the paper. I imported using @source jumpgarch.src but nothing works. I know I am missing something but not sure what exactly it is! When I use file--open--selected the source file, it just opened the source file. I also tried @jumpgarch.src and source jumpgarch.src but the results the same! Thanks in advance.
Prashant

[prashantj]

I think I have figured it out. But it is taking so long to estimate the model and the RATS still processing to replicate results of Chan and Maheu. Is it because of massive data or my RATS?

[TomDoan]

The jump GARCH example (with the Chan and Maheu data) should take somewhere around 3-5 minutes depending on the speed of your computer. They have 15000 data points, but split the models into subranges, so it's three separate estimates with anywhere between 3000 and 7000 data points. Note that they may have chosen to split the range because the model couldn't cope with the full 15000. If the model is taking an amount of time that doesn't seem to be (roughly) linear in the number of data points, you should put a TRACE option on the MAXIMIZE instruction to see what's going on with it. You can also try using only a portion of the range to see if there are signs in that that the model isn't working properly.

[prashantj]

Hi Tom,
You are right , it took 5 -6 min with "jumpgarch.rpf" file. I do not know what happened last time. My computer is 16GB RAM and I7 processor with 8th generation. You mentioned about using TRACE on Maximize. How do I do it? Where do I place TRACE? e.g. in following command.
maximize(parmset=meanparms+garchparms+jumpparms+poissonparms,pmethod=simplex,piters=2,method=bfgs,title="Constant Intensity Jump GARCH, 1928-1950") logl gstart end1950

[TomDoan]

It's an option. MAXIMIZE(TRACE,the rest

[prashantj]

Is it possible in RATS to examine spillover using Jump GARCH or estimate Multivariate Jump GARCH like that?

[TomDoan]

Obviously not with a univariate model. There is at least some literature on multivariate jump-GARCH models (see https://estima.com/forum/viewtopic.php?p=4474#p4474), but that doesn't strike me as very convincing. A bivariate process with independently firing Poisson jump processes in each series seems highly unrealistic, but making the jump processes dependent (particularly serially dependent, which is what you would think is natural) is probably quite difficult.

[prashantj]

Hi Tom,

I want to modify the Chan and Maheu (2002) method to ARJI-Exponential-GARCH model and apply it to my study. How do I introduce EGARCH in the program? What is the code for it? please

Your help is appreciated.

Prashant

[TomDoan]

Hi Tom,

I want to modify the Chan and Maheu (2002) method to ARJI-Exponential-GARCH model and apply it to my study. How do I introduce EGARCH in the program? What is the code for it? please

Your help is appreciated.

Prashant

Do you have a reference for ARJI-EGARCH?