There are a lot of factors that need to be considered, but I think 3 years should give a meaningful mean. You want to take into account what the player was projected to do. I know that projections aren't the end all be all, but they factor in past performance, along with other things like the player's age (whether the player is entering his prime or declining).
IMO, you can then get a fairly good idea about whether the player underperformed or overperformed, and whether regression is likely or a continuing trend in a certain direction is more likely.
I've posted this before: In terms of predicting howa player will play the remainder of the season, projections trump in season performance to date at any point in the season.