Unless it's a lattitudinal trend across a large number of figures being calculated at the same time, all being derived from the same initial figure using the same formula, the sum of which consistantly exceeds the original figure due to a bias in the method of calculation.
The point of the round to even method, if properly implemented, is that it recognizes the situations where error is likely to occur and neutralizes the tendancy to err in either direction.
What you said about avoiding rounding- ie transformation of the source figure- is true. But it isn't necessarily a longitudinal problem.