Video hashing has attracted increasing attention in the field of video searching. However, there was no technical research on the prediction of hash length, which is extremely important in mobile circumstance. In this paper, a hash length prediction method is proposed for video hashing in the case of video copy detection. The video feature is mapped to video hashes with different lengths via kernel-based supervised hashing (KSH). A part of the dataset is used as the training data to establish the relationship between the probability of collision (PoC) and hash length according to the probability distributions of the bit error rate (BER) of hash for non-copy and copy videos, respectively. Referring to this relationship, the approximate shortest hash length with approximate best performance for the whole dataset is predicted. Simulations demonstrate that the proposed hash length prediction method can estimate the approximate optimal length for the corresponding video hash, which can be used as a reference for the whole dataset.