Causal Discovery in Hawkes Processes by Minimum Description Length

Hawkes processes are a special class of temporal point which exhibit natural notion causality, as occurrence events in the past may increase probability future. Discovery underlying influence network among dimensions multi-dimensional is high importance disciplines where high-frequency data to model, e.g. financial or seismological data. This paper approaches problem learning Granger-causal processes. We formulate this model selection task we follow minimum description length (MDL) principle. Moreover, propose general algorithm for MDL-based inference using Monte-Carlo method and use it our causal discovery problem. compare with state-of-the-art baseline methods on synthetic real-world The experiments demonstrate superiority graph compared respect size results G-7 bonds price consistent experts’ knowledge.