Users’ behaviors on the Web shift over time in terms of pages of interest to users, queries issued to a search engine, and the underlying informational goals behind the queries. We examine how to model and predict user behavior on the Web over time. We explore several modeling approaches and develop a temporal modeling framework adapted from physics and signal processing that can be used for predicting user behavior. We also explore other dynamics of the Web, such as the detection of periodicities and surprises. We develop a learning procedure that can be used to construct models of users’ behaviors based on features of present and historical behaviors. The results of experiments indicate that, by learning how to predict user behavior using our framework, we can achieve significant improvements in prediction and explicitly learn when to apply each of a set of behavior prediction models.