It's going to be hard to pick one of the five finalists. But if the criteria remain (substantially) the same, I think the field may be narrowed significantly. I'm making one very crucial assumption here, of course -- that to the extent it is knowable, all five finalists (Rijndael, MARS, RC6, Serpent, and Twofish) will be equally secure. In that case, performance and confidence become major criteria. NIST marked down MARS and RC6 for their bias towards 32-bit platforms with particular architectural characteristics. RC6 is denigrated for a (relatively) low security margin; MARS is criticized for complexity. Serpent, though quite strong, is slow. Twofish is flexible, but perhaps too complex. Nothing negative was said about Rijndael in the summary -- it seems to be very secure, have a fast key setup time, and excellent performance on all platforms. When I look at those judgments (all taken from 2.7.3 of the NIST report), I suspect that MARS, RC6, and Serpent are going to be dropped for performance reasons. Twofish and Rijndael are both excellent performers across the board. The latter is simpler; the former seems to have a higher security margin (if I'm not reading too much into the difference between a "large security margin" and a "good security margin"). The answer may depend on the weighting of those two criteria.
