EventFlow: Forecasting Continuous-Time Event Data with Flow Matching

Continuous-time event sequences, in which events occur at irregular intervals, are ubiquitous across a wide range of industrial and scientific domains. The contemporary modeling paradigm is to treat such data as realizations of a temporal point process, and in machine learning it is common to model temporal point processes in an autoregressive fashion using a neural network. While autoregressive models are successful in predicting the time of a single subsequent event, their performance can be unsatisfactory in forecasting longer horizons due to cascading errors. We propose EventFlow, a non-autoregressive generative model for temporal point processes. Our model builds on the flow matching framework in order to directly learn joint distributions over event times, side-stepping the autoregressive process. EventFlow is likelihood-free, easy to implement and sample from, and either matches or surpasses the performance of state-of-the-art models in both unconditional and conditional generation tasks on a set of standard benchmarks. Many stochastic processes, ranging from consumer behavior (Hernandez et al., 2017) to the occurrence of earthquakes (Ogata, 1998), are best understood as a sequence of discrete events which occur at random times. Any observed event sequence, consisting of one or more event times, may be viewed as a draw from a temporal point process (TPP) (Daley & Vere-Jones, 2003) which characterizes the distribution over such sequences. Given a collection of observed event sequences, faithfully modeling the underlying TPP is critical in both understanding and forecasting the phenomenon of interest. While multiple different parametric TPP models have been proposed (Hawkes, 1971; Isham & Westcott, 1979), their limited flexibility limits their application when modeling complex real-world sequences. This has motivated the use of neural networks (Du et al., 2016; Mei & Eisner, 2017) in modeling TPPs.