bjautofit

[lnw]

I notice that in bjautofit AIC=-2ln+2*(p+q) but I thought it would have intercept in some cases, then %nreg would be better?

Otherwise sometimes, the manual calculated AIC may not match those from bjautofit.

[TomDoan]

I notice that in bjautofit AIC=-2ln+2*(p+q) but I thought it would have intercept in some cases, then %nreg would be better?

Otherwise sometimes, the manual calculated AIC may not match those from bjautofit.

For comparing models, you don't need to include parameters that are common to all. For instance, most descriptions of AIC for a regression model (or ARMA model) don't count the variance even though it is also an estimated parameter. However, if you compare a regression with homoscedastic errors with the same regression model with ARCH errors, you would want to include the variance in the count for the first regression to compare the two.