Trajectory prediction is a fundamental and challenging task for numerous applications, such as autonomous driving and intelligent robots. Currently, most of existing work treat the pedestrian trajectory as a series of fixed two-dimensional coordinates. However, in real scenarios, the trajectory often exhibits randomness, and has its own probability distribution. Inspired by this observed fact, also considering other movement characteristics of pedestrians, we propose one simple and intuitive movement description, probability trajectory, which maps the coordinate points of pedestrian trajectory into two-dimensional Gaussian distribution in images. Based on this unique description, we develop one novel trajectory prediction method, called social probability. The method combines the new probability trajectory and powerful convolution recurrent neural networks together. Both the input and output of our method are probability trajectories, which provide the recurrent neural network with sufficient spatial and random information of moving pedestrians. And the social probability extracts spatio-temporal features directly on the new movement description to generate robust and accurate predicted results. The experiments on public benchmark datasets show the effectiveness of the proposed method.