Recently, the number of Chinese films and audience as well as film market have been increased rapidly, while various box office prediction approaches have been proposed and investigated. However, these approaches only utilized stable information of films for prediction without using any dynamic information, making them unable to adjust predictions in real-time. Meanwhile, current methods only predict total box office of films but ignore everyday’s revenue and other audience information such as attendance rate, the number of audiences. In order to accurately predict each day’s audience data during the screening period, this study proposes a prediction algorithm based on Generative Adversarial Networks (GAN). For a film, this algorithm firstly converts the time-series audience data of available dates to frequency domain using Fourier transform, allowing global feature to be extracted. Then, a novel dynamic encoding method is proposed to obtain the dynamic of recent days. Finally, the deep spectral feature is extracted and future audience data is predicted using GAN. The experimental results show that the proposed algorithm can accurately predict the daily audience data of films. Compared to only using stable information, the proposed dynamic method can improve the prediction accuracy. When combined both stable and dynamic information, the best results are achieved. Therefore, the algorithm can provide useful information for the film marketing industry.