Diebold Mariano Test with the dmariano package

[Matthias Zielinski]

Hello,

I'd like to compare different forecasts with the Diebold-Mariano Test. There's a package "DMARIANO" created by Mr F. Baum. Does anybody know what number to use for the maxlag option? My Forecasts have a sample size between 15 and 50 observations. Diebold and Mariano suggest in their paper (1995) a lag of S(T)=k-1, because k-stepahead forecasts are at most (k-1) dependent. Does that mean that for a 1-stepahead forecast I choose maxlag=0? I'm not sure because the results I get for this are rather conservative. If anybody could help out here it would be greatly appreciated.

Best Regards,
Matthias Zielinski

[Carlo Lazzaro]

Matthias:
welcome to the list.
Please describe where does DMARIANO package come from (package dmariano from http://fmwww.bc.edu/RePEc/bocode/d). The reason for providing this detail are covered in FAQ. Thanks.
As far as your main question is concerned, did you take a look at the help file that comes with the abovementioned package?
Have you already googled yourself about this issue?

[Matthias Zielinski]

Thanks for your reply Carlo.
Yes, it's the package from http://fmwww.bc.edu/RePEc/bocode/d, created by Mr Baum from Boston College. Yes, I looked into the helpfile and it says that it uses the Schwert Criterion as default, which is 12*(T/100)^1/4 with T being the number of observations, which suggests lags between 7 and 10 for the upper stated case. I googled it and found often the suggestions made by Diebold and Mariano in their paper "Comparing Predictive Accuracy" from 1995, which is (k-1) because of the above mentioned reason. I want to compare a VAR-model forecast to a benchmark AR-forecast by using this test among others. The test statistic and thus the test decision reacts very sensitive to the chosen maxlag option, so it irritates me a little that there's such a gap between the literature suggestion and the software default. I also contacted Mr Baum personally, but wanted to post in this forum as well in case anybody can help in this regard. I looked into the code but it is just too advanced for me to understand it fully. When I use the literature suggestion, I get rather conservative results compared to similar setups in different papers, so I wanted to make sure there's no misunderstanding on my part.

Kind regards,
Matthias