The automatic detection of rumors in social networks has gained considerable attention from researchers and practitioners during the last decade, due to the consequences of the spread of disinformation in public opinion. Most of the existing methods make use of features extracted from conversational threads, user profiles, and structural information of the network. These features are difficult to capture in practice and are often only partially available during the spread of rumors. In this paper, we study an unexplored approach in rumor detection: time series classification (TSC). By modeling the problem using time series, we avoid using lexical or structural characteristics of the network. Instead, we use information that is simpler to capture, such as the volume of tweets and the number of followers and followees of the users involved in a story. In this way, the characterization of the story is not related to specific users, but to variables aggregated at the event level.We introduce a TSC-based model for detecting rumors based on hypergraph partitioning, aligning time series prototypes with rumor classes. Our approach uses a Siamese network to train a rumor detection model in a supervised way, minimizing the distance between the time series of the training examples and the prototypes of their class. Results on benchmark data show that our approach surpasses other TSC-based methods in detecting rumors. Also, we compare our methods performance with methods that make use of lexical and structural characteristics. Our experiments show that our method has advantages in time-sensitive contexts, outperforming the state of the art in early detection scenarios with incomplete information.