When predicting sequences of events over time, simpler models like Hawkes processes were thought to be worse than complex neural networks. But earlier comparisons may have been unfair because they used overly restricted versions of the simpler models.
We systematically compared Hawkes processes with flexible, learnable components against neural temporal point process (TPP) models on event sequence prediction tasks, including large medical datasets.
The performance gap largely disappears when Hawkes processes are given flexible parts. The advantage of neural models is mostly in predicting when events happen, not which events. This means simpler classical models are often sufficient, guiding model choice.