Travelling in metropolis is becoming increasingly difficult for people due to the growing population and busier traffic. Since taxi is important public traffic tools in cities, knowing how long it will take to find a taxi can be helpful for customers to plan their schedule and choose the best place to wait. This paper presents a method to predict the waiting time for a customer at a given time and location from historical taxi trajectories. With a parametric and a non-parametric model for arriving flow of vacant taxis, it predicts the time a single passenger needs to wait for a taxi service. Emulation experiment with large-scale taxi dataset in Hangzhou was conducted to verify the new method. The evaluation results validate that the mean prediction errors for both the parametric and non-parametric model are about 4.5 minutes. Moreover, the probability of that the prediction error is below 5 minutes for parametric model is about 83 percent.