With the development of the Internet and the continuous expansion of the online video market scale, it is increasingly important to accurately predict broadcast volume before the launch of videos. On the one hand, it can provide investors and producers with recommendation scheme for video shooting. On the other hand, it can fully understand users' preferences and find videos which potentially to be popular. In allusion to problems of low predictive accuracy and lack of practical application value in video prediction research, this paper, based on the related data of online video website TED, obtains a prediction model with high-precision broadcast volume through feature selection and model fusion. Further, it analyzes the impact of video themes, the number of languages and official events, which can provide some reference for investors and producers of different scales before the videos go online.