For what it's worth, there is a method called "bucket rounding" that is often used in transportation planning (but nowhere else according to google) which can help in this type of problem (within the limits of machine precision). The process is very simple: the "residue" from each rounding operation in a sequence is retained to be applied as a bias to the next one. So for merlyn's example:

#   Value   Residue    Rounded     Comment
1    10.3     0.0        10        0.3 is added to the residue
2    10.3     0.3        11        10.3 + 0.3 = 10.6, which rounds to 
+11 with -0.4 residue
3    10.4    -0.4        10        
     ====               ====
     31.0                31
[download]

This is most commonly applied in the factoring of 2D matrices representing trips from location to location. While the value of each cell is not usually affected by a minor rounding error, the total of the matrix needs to be held constant.

I'm certainly not qualified to say whether this would be appropriate for financial work though.

--
I'd like to be able to assign to an luser