question about Beveridge-Nelson and Unobserved-Components De

[hardmann]

Dear Tom:

I repeat the example of Morley, Nelson & Zivot(2003), "Why Are the Beveridge-Nelson and Unobserved-Components Decompositions of GDP So
Different? When I substitute Chinese quarterly real GDP seasonally adjusted and slightly alter origin code, I encounter some problem. When I estimate UC-UR, I get rho 1.25, which is rather positive than negative and more than 1. I am not sure that before estimating BN model, the ARIMA(p,q) form must be specified. When I use rgdp data adjusted by X12, I get ARIMA(2,1,2). According to this specification, my result is different form standard procedure @bndecomp. I do not know what is up?

Attached is my data. I plea you examine my data if possible.

Best regards.

hardmann

[TomDoan]

You have a rather short data set for this type of analysis. MNZ had 50 years; you have 16. The first difference of GDP is fairly close to white noise so the estimates of ARMA(2,2) models aren't very precise---if you change your method you get very different estimates. If you put a PRINT on the BNDECOMP procedure (that's why it's there!!) you'll see that it doesn't converge on an ARMA(2,2) model with the set of options used by the procedure (which has the series de-meaned first). De-meaning the series first versus estimating the mean as part of BOXJENK should give almost identical results in large samples, but this is small enough that they don't.

[hardmann]

Thanks Tom.