The book "Real-Time Collision Detection" suggests that the AABB centroid is a decent approximation to the bounding sphere. In fact, the AABB centroid is much more accurate than the center of mass of the points, in general.
A quick survey of the techniques mentioned in the book (Chapter 4):
- Centroid of AABB: quick and dirty, not terrible
- Center of data in direction of maximum variance (eigenvalue/eigenvector of covariance matrix) --> This method seems like it could generalize to find a good bounding ellipsoid of your data, potentially better for high aspect-ratio point sets
- Randomized O(N) algorithm (Welzl's algorithm) which maintains a set of points which must lie on the minimum sphere boundary. An implementation is available in CGAL.
The book is written toward video game programmers, so there is an emphasis on being "good enough" quickly rather than necessarily correct. A typical use-case for these bounding volumes in this context is for collision detection. Since the goal is to minimize the overall time to do the collision detection, having exact bounding volumes that are potentially expensive to compute is counter-productive, hence the emphasis on approximate methods.
Here's an official link to the book, but there are other "less official" links floating around the internet as well...