This work addresses the online detection of sub-events using Twitter stream data. We formulate the process of sub-event identification as an outlier detection problem using three statistical methods: Kalman Filter, Gaussian Process, and Probabilistic Principal Component Analysis. These methods are used to construct the probability distribution of percentage change in the number of tweets. Outliers are identified as future observations that do not fit these predicted probability distributions. Five real-world case studies are investigated to test the effectiveness of the methods. Finally, we discuss the limitations of the proposed frame-work and provide future directions for improvement.