We investigate the accuracy of deterministic, regional-scale ground motion simulations of moderate magnitude earthquakes in southern California, and the influence that the models involved have on synthetic results when compared to data. Nowadays, advances in earthquake ground motion simulation algorithms and models, and growth of high performance computing systems and applications facilitate regional-scale simulations of earthquake ground motion using numerical approaches. However, before simulations can be used in engineering practice, much work is needed to confirm the robustness of models and simulation methods. This requires a continuous effort on simulation validation through comparisons with data. We evaluate the accuracy of simulations using quantitative metrics of physical meaning to both seismologists and engineers. We show that validation results are significantly controlled by the choice of the models, including crustal velocity models, attenuation models and parameters, and source models. We focus our attention on the selection of the appropriate velocity model by performing a large set of simulations for multiple past events in southern California, using the different community velocity models available for this region. After identifying the model that consistently yields the best possible approximations, we investigate the influence of different attenuation viscoelastic models, and that of the definition of attenuation parameters and quality factors, as well as the relative influence of point versus extended source models. We concentrate on moderate events in the greater Los Angeles basin area for which there are significant number of high-quality data. We analyze the results through quantitative goodness- of-fit measures and shed light on the relative weight of these factors with respect to each other, and how they influence validation results, and thus simulations as a whole.