[Gretl-devel] svars

Hi, with respect to analyzing SVARs I've taken a relatively close look at the (AFAIK quite recent) 'vars' package in R, and also read Jack's doc on his preliminary gretl svar functions (well I also looked a little at the functions themselves). First of all, R's 'vars' package is quite complete and nicely documented. So my first reaction is: don't bother with reinventing the wheel yet again and point interested users to that package. I could also write some wrapper functions (packages) to access them from the gretl side. Two possible caveats: 1. Bootstrapping the impulse responses appears to be very slow in R-vars. I compared (roughly) the same setup in Jmulti and in R-vars, and Jmulti (Gauss in the background) was at least 10 times faster than R-vars. (Admittedly I ran Jmulti on Windows on a different computer, but the hardware specs are not too different and 9 seconds vs. 3 minutes is drastic enough for me.) 2. R-vars relies on R-urca from the same author to specify the VECM (for SVEC modeling), and AFAICS in R-urca it's not possible to place such elaborate restrictions on the cointegration relations as gretl (remember the Boswijk/Doornik paper) or Jmulti can do. @1: In general I would tend to think it's better to help R-vars get faster instead of developing a competing implementation. (Better to collaborate than to multiply the needed efforts.) But I don't know if that's feasible, I really don't know much about R or about the implementation in R-vars. Also, since Jack has already done quite an amount of work on this, gretl wouldn't start from scratch, so maybe there wouldn't be that much duplicate effort (not counting the sunk effort, of course). @2: For me this seems a more serious issue, but that's of course driven by my personal preferences. In my view modeling the cointegration space properly for non-trivial setups often requires placing restrictions on the cointegration space. So the fact that this feature is missing from R-vars could be a reason to do a gretl-native implementation. (Not sure wether native would necessarily mean c-coding though, pure gretl may be enough?) OTOH, it could be argued that this is very specialized and people should stick to Jmulti (or Anders Warne's SVAR program which actually does much more than that). Jack, AFAICS your functions also do not allow to do restricted VECMS, right? Comments? thanks, sven