Most event generators are wonderful for some things but have significant limitations for others. Some are very easy to use and convenient to run, but will only do 2 → 2 processes in the main scattering event (which leaves out many important 2 → 3 and 2 → 4 processes that can be important standard model backgrounds to new physics signals; for example g g → t + tbar + Z is a source of large missing energy and leptons.) Many cannot handle the cascade decays of new particles correctly; they may fail completely (because the phase space integrals required simply take too long) or they may simply discard some important information (such as the correlations between the spins of the new particles and how those correlations propagate into the decay products.) Some generators that can handle these issues pretty well are unfortunately harder to modify to accomodate new-physics processes.

There is no simple solution here — it is necessary to understand both the generator you are using and the physical processes (signal and background) that you are simulating, in order to avoid very significant errors.

Moreover, even if your event generator correctly computes tree level amplitudes, this doesn’t mean it does the physics right. Loop corrections are huge in QCD (more precisely, without a loop correction, tree amplitudes suffer from large ambiguities, since they are proportional to a power of a running coupling, whose value is not determined at tree level!) This can be very roughly dealt with, process by process, by normalizing the rate for each process using a next-to-leading order computation of that rate and hoping the tree-level result is still giving the correct kinematic distributions. But this is not practical for simulating many processes at once, since typically event generators are not written in such a way that you can easily adjust the normalization of each process by hand. One should also remember that parton distribution functions are needed for predicting the rate of any given process, but these functions are neither perfectly determined from experiments (especially gluon and heavy quark distributions, which are important at LHC) nor free from effects of loop corrections. So don’t take any one of our black box data files too seriously — our simulation of the signal from a new physics model is not, for these and other important reasons, what would actually be seen at LHC if this model were a correct description of the real world. The errors are very hard to quantify without a detailed study of both the signal and standard model backgrounds.

Another modern effort in event generation involves the MC@NLO project (Monte Carlo at Next-to-Leading Order).

Discussion of Pythia vs. Herwig vs. ...