Different models are used in literature to approximate the complex dynamics of a humanoid robot. Many models use strongly varying model assumptions that neglect the influence of feet, discontinuous ground impact, internal dynamics and coupling between the 3D coronal and sagittal plane dynamics. Often motivation is lacking why a certain model should be chosen or why certain dynamical aspects may or may not be neglected. A simple model like a linear inverted pendulum with low computational cost is frequently used on-line in foot placement algorithms, where model inaccuracy may lead to low control performance. A full 3D multi-body model with higher computational cost is typically used off-line in a simulator, where model inaccuracy may lead to unreliable results. This paper contributes experimental validation of both types of models by comparing them to the real dynamics of the humanoid robot TUlip during balance recovery side-stepping. Hereto, we evaluate the aforementioned most common model assumptions. The validated multi-body model that we implemented in the ROS Gazebo simulator is released on-line.