I fit the shadow projection to the view frustum, and do not clip against the scene bounds, because in Common Techniques to Improve Shadow Depth Maps it is stated:
It is also possible to clip the frustum to the scene AABB to get a tighter bound. This is not advised in all cases because this can change the size of the light camera's projection from frame to frame. Many techniques, such as those described in the section Moving the Light Texel-Sized Increments, give better results when the size of the light's projection remains constant in every frame.
However, in the paper from the Nvidia SDK the frustum is further clipped against the scene bounds to achieve a maximum used area for the shadow map.
This confuses me. Did the guys from MSDN miss something, is the Nvidia paper simply older or am I not understanding the Nvidia approach correctly?
Edit: I was right in not understanding the Nvidia approach, basically not at all. I confused the crop matrix with a clipping against scene bounds.