FinRipple: Aligning Large Language Models with Financial Market for Event Ripple Effect Awareness

Yuanjian Xu, Jianing Hao, Kunsheng Tang, Jingnan Chen, Anxian Liu, Peng LIU, Guang Zhang

Keywords: Large Language Models, Time-Varying Knowledge Graph, Event Studies, Asset Pricing

TL;DR: We propose FinRipple, a novel training framework that enhances the market analysis capabilities of large language models

Abstract: Event studies have been fundamental in finance, focusing on analyzing the ripple effects of sudden market events. Accurately predicting these effects is crucial for informed decision-making and effective risk management. However, the dynamic complexity of financial markets and the lack of unified modeling tools make this task challenging. Previous models, constrained by simplistic assumptions and limited scopes, have struggled to address this complexity effectively. In contrast, large language models (LLMs), with their emergent reasoning abilities, offer a promising solution. In this paper, we introduce $\textbf{FinRipple}$, a novel training framework that enables LLMs to align with market behavior and develop the capability to analyze the ripple effects of sudden events. We first construct a time-varying financial knowledge graph (KG) that is both financially meaningful and noise-reduced to accurately represent the market state. These KGs are then integrated into the LLM using adapters as memory modules. Additionally, we align the LLM with market dynamics by integrating FinRipple with classic asset pricing theories through a reinforcement learning framework. This market-alignment process collects feedback that enhances the LLM's foundational ability to analyze financial events and explain market anomalies that traditional models fail to address. Our key contributions are as follows: (1) We are the first to define the underexplored task of ``event impact prediction''. Our framework not only establishes this task but also provides an open-source benchmark, creating a unified evaluation standard for both academia and industry; (2) FinRipple complements classic asset pricing models by combining strong theoretical foundations with AI-driven capabilities, offering an enhanced analysis of residuals unexplained by traditional models. We also demonstrate its potential for practical applications such as portfolio management; (3) We conduct a comprehensive analysis to ensure that the results generated by LLMs in our framework are more logically consistent and credible, thus improving the reliability of insights for financial decision-making.

Primary Area: foundation or frontier models, including LLMs

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 3612