Is more data better?

Conventional wisdom usually suggests that agents should use all the data they have to make the best possible prediction. In this paper, it is shown that agents may make better predictions by discarding old data if their model is mis-specified. The applicability of the results to some economic models is also demonstrated.