Latent Diffusion for Event Driven Asset Pricing

ACL ARR 2025 February Submission7722 Authors

Abstract: Event studies have garnered widespread attention in academic research due to their significant impact on asset prices, playing a crucial role in both risk management and the understanding of market dynamics. However, existing methods face notable challenges. One such issue is the lack of effective multimodal alignment schemes, with many approaches relying on discretizing time series data when aligning it with language modalities. This often leads to the loss of valuable information, as continuous frameworks are better suited for capturing the dynamic nature of market behavior and accurately tracking rapid shifts in asset prices, as demonstrated by extensive theoretical and empirical work. Additionally, these methods struggle to model the inherent randomness of financial systems. To address these challenges, we introduce a Multimodal Latent Diffusion model specifically designed for event-driven asset pricing. Our approach integrates textual representations of sudden events with financial time series in a continuous latent space, preserving subtle temporal variations and fully leveraging the rich semantic cues embedded in the event-related text. Through comprehensive experiments and case studies, we demonstrate that our method consistently enhances predictive accuracy for event-driven asset pricing, while also expanding practical applications for risk management.

Paper Type: Long

Research Area: NLP Applications

Research Area Keywords: NLP Applications, fnancial/business NLP

Contribution Types: Model analysis & interpretability, NLP engineering experiment, Publicly available software and/or pre-trained models, Data resources

Languages Studied: English

Submission Number: 7722