[R-sig-finance] VAR, VECM, Kalman, ... non-R software recommendations?

[Dirk Eddelbuettel]

Patrick,

Thanks for yet another very helpful post in this thread!

On Fri, Aug 20, 2004 at 03:29:24PM -0500, Patrick Brandt wrote:
> I've been a RATS user for about 6 years (*NIX and Windows) and a Stata 
> user for 10.   RATS is a great package for doing all of the standard 
> econometric time series, esp VARs and VECMs.  The good thing is that it

Yes, I used it a for little bit a long time ago.

> includes pre-packaged routines for doing impulse responses, 
> forecasting, and decompositions.  It is my favorite for time series, 
> because it is one of the few packages that does not strive to do 
> everything -- it works to do time series well.

That seems to be a consensus view.

> For basic multivariate time series modelling, RATS and Stata can do the 
> job well.  I have my quibbles with both (such as Stata not having a 
> well defined set of time series "objects" or methods that really 
> understand how to work with ts data) and RATS "unique" syntax.  Both 
> will allow you to do the standard VAR and VECM models in Hamilton or 
> Johansen.
> 
> That said, programming in RATS is not for the faint at heart.  Using 
> the standard routines works well, but once you start doing more exotic 
> things (complex, high dimensional SVARs come to mind), or posterior 
> simulations for Bayesian VARs (BVARs), things get more complicated (in 
> part because the RATS syntax has a combination of old fashioned Fortran 
> and C declarations).  My guess is that while these can all be done in 
> RATS with some degree of effort, the effort necessary to do them in 
> Stata will be monumental.  I find that whenever I work in RATS I have 
> to have a set of manuals nearby.
> 
> For these reasons (and as part of a larger project to model 
> international conflict data and political economy data), I have started 
> on an R package that will estimate VARs, Bayesian VARs, and 
> Markov-switching BVARs.  This is being done in R for the obvious 
> reasons: 1) it is free / open source, 2) R is gaining wider use in the 
> social sciences, 3) I can write the computationally intensive functions 
> for the BVARs and MS-BVARs in C++ and make them very fast, and 4) the 
> object, scope and method aspects of R lend themselves more easily to 
> programming these models.

That sounds very intriguing too, and would complement the kalman filter code
in R's base, as well as Bernhard's urca package.  

Any expected timelines?

> At present, few if any of these VAR / SVAR extensions are present in 
> Stata (even with the new VAR routines they have added, one cannot 
> estimate the BVAR models, or any error bands for the impulse 
> responses).  RATS has the capacity to offer all of these methods.
> 
> Another option is Ox, which is open for academic use, and a reasonable 
> fee for non-academic use: http://www.nuff.ox.ac.uk/Users/Doornik/  Ox

Yes, I also looked at Ox, back when it came out and every now and then
afterwards. I find its licensing to be the most annoying -- free for you but
not for me. Weird hybrid.  That said, a possible contender in this too.

> can estimate VAR and VECM models, with many specialized addons.  Also, 
> the Ox syntax is remarkably similar to C++.  There is a package for Ox 
> that will do most of the state-space modeling outlined in Durbin and 
> Koopman as well.

>From what Koopman said, it is actually the same codebase he contributed to
Ox, gave to Brian Ripley for R around release 1.5.0.  Zivot and Wang credit
it explicitly in the Finmetrics book.

Regards, Dirk 

-- 
Those are my principles, and if you don't like them... well, I have others.
                                                -- Groucho Marx