An apparatus and method for motion vector prediction for a current block in a picture are disclosed. In video coding systems, the spatial and temporal redundancy is exploited using spatial and temporal prediction to reduce the information to be transmitted. Motion Vector Prediction (MVP) has been used to further conserve the bitrate associated with motion vector. In conventional temporal MVP, the predictor is often based on a single candidate such as the co-located motion vector in the previous frame/picture. If the co-located motion vector in the previous frame/picture does not exist, the predictor for the current block is not available. A technique for improved MVP is disclosed where the MVP utilized multiple candidates based on co-located motion vectors from future and/or past reference pictures. The candidates are arranged according to priority order to provide better availability of MVP and also to provide more accurate prediction. Furthermore, the MVP technique disclosed can be operated in a closed-loop fashion so that no additional side information or minimum additional side information is required.