In this paper, we propose and evaluate the accuracy of a new traffic model that predicts, in the short term, the size of the B-frames of MPEG-4 AVC compressed video. To quantify the prediction accuracy, a model for the prediction error is also proposed. In a broadband cable network, the prediction model can be employed by a cable headend to provision video bandwidth efficiently or more importantly, reduce bit rate variability and bandwidth requirements via selective B-frame dropping, thereby minimizing buffering requirements and packet losses at the set top box. It will be shown that the model provides highly accurate prediction for a wide variety of video types, in particular, high action and regular movies encoded in high quality resolution. This is an important benefit since high motion MPEG videos demand the highest bandwidth and generate the widest range of bit rate requirements. Although the prediction model is very simple to implement, this paper also proposes two methods to simplify its operation further without sacrificing prediction accuracy.