Kim et al.(2003) studied a prediction model for students likely to transfer. In their study they claim that a training data construction scheme is better than other schemes, which trains neural network on the data from the year right before prediction year. One problem with their claim is that it is based on rather high prediction error rate. In this paper we establish a more sound comparison for various training data construction schemes and check validity of their claim. It turns out that the favored scheme has sufficient advantages over other schemes.