This seems like an optimization question. As such the answer, as always is to profile it.
If generating the matrices doesn't take too much time on the fly, then it is simply easier to do on the fly, as you don't have to write and export/import of matrices, and add it to your file format.
And perhaps you could simply use other file formats natively and not have to bake your resources.
On the other hand if your application is very very animation intensive, it may be worth it to move the matrix calculation to the baking stage.(Tough I would seriously doubt that.)
In the end this looks like a premature optimization to me. If you are interested, then you should profile it first, and if the results say you spend too much time calculating matrices, then do it.
PS: If you do profile it, don't forget to post the results.
EDIT:
To answer your comments, there is no reason to precalculate the matrices. But I see no problem in doing so. But if you also want to interpolate... It doesn't look straightforward. Maybe you could store a matrix for every frame but that would be quite a lot of data.
Well To expand on this here is some calculation:
Assuming 4x4 matrices of float(4byte) data.
Lets assume you have 16 bones in your model.
And you want 200 animation frames. And you want interpolation so you do 2000 animation frames instead(10 subframes)(Assuming a bad case scenario).
This would require 2000kB of memory/disk space.
Thats not a lot.(Except if you are doing skinning on GPU, then if you consider 100 different models, this is 200MB animation data, which is quite significant.) But you could just easily calculate it on the fly with even more accuracy, and don't need to have larger files.
And mainly: write, test and maintain the extra code to do it the other way around.
So the more I think about it the less feasible reasonable it seems.
