When RNN-based Marked Point Processes Fail in Real-World Finance: A Tiny Paper

ICLR 2025 Workshop ICBINB Submission17 Authors

Track: tiny / short paper (up to 2 pages)

Keywords: recurrent neural networks, time series, finance

Abstract: Neural Marked Temporal Point Process (MTPP) models have shown promise in controlled benchmarks for forecasting and event pattern modeling in finance. However, when deploying Recurrent Neural Network (RNN)-based MTPPs on large-scale, high-dimensional financial event streams, we encountered unexpected challenges: ballooning parameter sizes, increased computational costs, and training instability. This short paper outlines (1) the financial use case, (2) the literature-proposed neural MTPP solution, (3) the negative outcomes observed, and (4) our investigation into why standard MTPPs fail to generalize as promised in real-world conditions.

Anonymization: This submission has been anonymized for double-blind review via the removal of identifying information such as names, affiliations, and identifying URLs.

Submission Number: 17