It sometimes happened to me to have 1 hour of render per image, but it is exceptional!
It is especially when there is a combination of reflections, transparencies and lights that there are many calculations, for the volumetric clouds too.
The particles require sometimes many calculations if they are not well parameterized.
I finish just a footage 6 seconds (25i/s), and that took 3:40 ' for the whole (maximum quality).
There is about 50 peoples who are jogging reflected in a car and a few through the glasses.
Can you explain how you replace the transparency by caustic, I would like to test?
With the .bvh, there is often work to make in the keyframes, that depends on the character whom you use.
You always select the hip for the importation and you select the existing skeleton in the import window.
On Carrarators, we have some links for  downloading some free .bvh.
I also started to make packs of .bvh converted into NLA (it's a beginning).