The explosive growth of multimedia applications renders the efficiency of network resource allocation a problem of major importance. The burstiness of video traffic, in particular, calls for traffic control solutions that will help prevent significant packet losses. Such losses can lead to unacceptable quality of service (QoS) and quality of experience (QoE) to users. In this paper, we focus on a large variety of H.264- and H.265-encoded video traces with different GoP patterns. Different versions of each trace, in low, medium, and high quality have been used in our study. We evaluate the accuracy of an existing video traffic prediction approach for the size of B-frames, and we propose a new Markovian model that predicts B-frames' sizes with significantly higher accuracy. B-frame size prediction can be used in order to reduce bandwidth requirements and smooth the encoded video stream, by selective B-frame dropping, when the model predicts larger upcoming B-frame traffic than the network can handle.