In this paper a new proposal for memory-based Collaborative Filtering algorithms is presented. In order to compute its recommendations, a first step in memory-based methods is to find the neighborhood for the active user. Typically, this process is carried out by considering a vector-based similarity measure over the users’ ratings. This paper presents a new similarity criteria between users that could be used to both neighborhood selection and prediction processes. This criteria is based on the idea that if a user was good predicting the past ratings for the active user, then his/her predictions will be also valid in the future. Thus, instead of considering a vector-based measure between given ratings, this paper shows that it is possible to consider a distance between the real ratings (given by the active user in the past) and the ones predicted by a candidate neighbor. This distance measures the quality of each candidate neighbor at predicting the past ratings. The best-N predictors will be selected as the neighborhood.