2

I have multiple financial time series data (FX-rates, commodity prices) that have been recorded daily (without weekends) for the past six years and want to analyze their effect/influence on the stock price of a certain company.

I have tried to do this via ARIMA/ARMAX but somehow this does not lead to any plausible result (especially the in-sample forecasts generated by these models are rather poor).

Someone has given me the hint that maybe GARCH is a better method of modeling the dependencies of the above mentioned variables. I am very new to econometrics and do not have a mathematical background. Therefore I am looking for a simple explanation on how to come up with such a multivariate GARCH model (most preferably in Gretl). I would need some sort of manual/tutorial that (1) avoids all the math that underlies GARCH as much as possible and (2) describes the process of choosing the different parameters ($p$,$q$), the necessary independent variables that need to be included in the model, etc., step by step (something like this ARIMA manual but for GARCH). So far I have only found very sophisticated scientific papers that were far too mathematically for me to grasp...

I would very much appreciate any help - or if you feel I am totally off track, I also welcome any kind of corrections!

Richard Hardy
  • 54,375
  • 10
  • 95
  • 219
Jonathan Rhein
  • 123
  • 7
  • You may start off by checking out the existing threads about order selection for GARCH models (although most of them consider univariate cases), perhaps [this](http://stats.stackexchange.com/questions/154754/how-to-choose-the-order-of-a-garch-model?rq=1), [this](http://stats.stackexchange.com/questions/77925/procedure-for-fitting-an-arma-garch-model?rq=1), [this](http://stats.stackexchange.com/questions/143517/arma-garch-estimation-process-in-practice?rq=1), [this](http://stats.stackexchange.com/questions/159807/arma-garch-model-selection-fit-evaluation?rq=1), and links therein. – Richard Hardy Jan 07 '16 at 12:36
  • And [here](https://cran.rstudio.com/web/packages/gets/) is a link to an R package that does automated model selection (both for conditional mean and conditional variance) given the regressors. It builds on reputable work of Jurgen Doornik ("Autometrics") and colleagues. – Richard Hardy Jan 07 '16 at 13:39
  • thank you Richard for your quick and very helpful answer! Actually I have never used R really, is there any such solution (as "gets) that works in GRETL or even as add in excel or some other statistical software? Also would there be a possibility of contacting you and showing you some of my data? Thank you! – Jonathan Rhein Jan 07 '16 at 15:46
  • Sorry, I am not ready to get my hands dirty now. I am unfamiliar with GRETL, so I have nothing to offer there; I also do not work with Excel's add ins. Also, I forgot to mention earlier: do not expect good fit for financial returns data (if you achieve that, you might have overfit) and watch out for spurious regressions if you use levels rather than returns. – Richard Hardy Jan 07 '16 at 16:24
  • I checked out Gretl's [manual](http://gretl.sourceforge.net/gretl-help/gretl-guide.pdf) and was not able to find multivariate conditional variance models; I only found univariate ARCH and GARCH (see p. 217-230). Should I look somewhere else, or can I conclude that one *cannot* fit multivariate conditional variance models in Gretl? (If so, your question becomes irrelevant as long as you insist on using Gretl.) – Richard Hardy Jan 13 '16 at 19:29
  • The package MGARCH is available. As of now, it only handles unrestricted BEKK models, though. – Jack Lucchetti Jun 09 '17 at 10:49

1 Answers1

1

I'm sorry to say that Richard's hunch is correct, multivariate GARCH models aren't in gretl yet, neither in the core nor in the "gig" (garch-in-gretl) add-on. (Such a feature can be added through a function package in gretl's scripting language hansl, leveraging the ML routines without having to do the coding in C, but so far nobody seems to have contributed that.)

Sven S.
  • 524
  • 2
  • 8
  • Just to update the information a little: Since April 2017 there has been a function package for gretl called MGARCH that "handles unrestricted BEKK models (with analytical derivatives)". (I am not the author.) – Sven S. Nov 21 '18 at 22:12