I used to run an online service predicting college admissions. As I would update the algorithms to incorporate new results, users' predictions would change. At first, I thought people would be pleased. Soon, I discovered that people really don't like the idea of their prediction changing when they have done nothing to cause it to change. Subsequently, I would lock in the values that had been "seen" by the user, and allow them to update their prediction, if they wanted to, when new models came online. This eased a lot of the users' expressed frustration.