Rationale for using empirical moments

Dear authors,

My quick question is related to the moments of a DSGE:

once I have estimated a standard DSGE model, I have the option to compare either the theoretical moments or the empirical moments (periods>0 in stoch simul).
What is the rationale for using either the theoretical or the empirical moments ?
Is there a sound reason why one should go for either case, given that they are different … or should the pragmatic reason prevail to use whichever fits best to the data ?

Many thanks in advance ?

Historically, people have often used simulated moments with the idea to capture e.g. small sample biases. But then they took the average over many replications (e.g. Hansen 1985), resulting in effectively a large T. This often results in moments very close to the theoretical ones (but with the advantage that you can put some uncertainty bands around the mean estimates). For linear models, the difference is often not that big and I therefore prefer theoretical moments. But tastes may differ.