The RATS Software Forum

Dear Tom:

I still replicate paper of Morley, Nelson & Zivot(2003), "Why Are the Beveridge-Nelson and Unobserved-Components Decompositions of GDP So Different?,"
I encounter a new question.
The boxjenk code used is "boxjenk(maxl,ar=2,diffs=1,ma=2,const) lgdp" .
Without maxl option, the result is different, so BN decomposition is also different. While I use boxjenk with maxl option to other dataset, the model does not converge, but default specification is ok. I also review reference manual but no answer.

Best regard.
Hardmann.

In the User's Guide, we note that the MAXL method is more sensitive to problems with initial conditions, particularly with models that may be difficult to fit as in this case.

Try using the conditional least squares estimates as initial values for the MAXL estimation. For example:

boxjenk(ar=2,diffs=1,ma=2,const) lgdp
compute coeffs = %beta
boxjenk(maxl,ar=2,diffs=1,ma=2,const,initial=coeffs) lgdp

Regards,
Tom Maycock
Estima