[R-SIG-Finance] auto.arima

[aschmid1]

I have a sample X with 2600 points and 18 external regressors (dummy 
variables). X is integrated according to KPSS test: auto.arima() gives 
(2,1,2) model with a few statistically significant coeff. for dummies. 
Then I run auto.arima for Y=diff(X) and... get (6,0,6) model with very 
different coeff for dummies. Shouldn't I get (2,0,2) model for Y or at 
least similar coeff for dummies? Also, arima(2,0,2) for Y is very 
different from  (2,1,2) model for X. Which model should I believe?
BTW, auto.arima with ADF (rather than KPSS) for X gives (2,0,3) model 
with ar1=1.8346 and ar2=-0.8351 but its coeff for external regressors 
are very close to those for (2,1,2) -and I'm interested primarily in 
them. Still such uncertainty in arma coeff troubles me and I'll greatly 
appreciate your comments.

Thanks! Alec