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

In one sentence

Summary

The paper addresses the limitation of existing event studies that fail to capture complex, cross-entity ripple effects in financial markets. It introduces FinRipple, a framework that integrates time-varying knowledge graphs (KGs) with large language models (LLMs) to model dynamic corporate relationships. The authors formalize ripple effect prediction as a standardized task, aiming to enhance LLM reasoning by grounding it in market structure and classical asset pricing theory. This approach overcomes the structural unawareness and hallucination issues common in direct LLM applications for finance. The methodology involves three stages: constructing time-varying KGs from unstructured data (news, patents, supply chains), injecting this knowledge into LLM adapters via instruction tuning, and aligning predictions with market reality using Proximal Policy Optimization (PPO). The reward function is derived from CAPM residuals, ensuring predicted shocks explain unexplained return variance. Experiments utilize 110,000 news articles and S&P 500 data, comparing FinRipple against RAG, zero-shot, and in-context learning baselines across multiple LLM architectures. Evaluation metrics include explanatory power (R2) on CAPM residuals, ANOVA significance, and refusal-to-answer rates. Results show FinRipple significantly improves explanatory power, with R2 values up to 0.59 on Fama-French 5-factor residuals, outperforming all baselines. The model demonstrates superior instruction following and lower refusal rates. In portfolio management, a strategy based on FinRipple predictions achieves a Sharpe ratio of 1.153 and lower maximum drawdown compared to equal-weight and Markowitz benchmarks. Limitations include the computational cost of training and the dependency on the quality of KG construction. The paper provides an open-source benchmark for ripple effect prediction.

Research problem

Financial markets exhibit complex dynamics where localized events trigger ripple effects across interconnected entities, which traditional static event studies fail to capture. While LLMs offer reasoning capabilities, they lack structural market awareness and struggle to analyze these dynamic ripple effects accurately. The paper addresses the need for a framework that aligns LLMs with financial market structures to predict these cascading impacts.

Core contributions

• Proposes FinRipple, a framework integrating time-varying knowledge graphs with LLMs to model dynamic market structures.
• Introduces a financial theory-guided reinforcement learning approach using CAPM residuals as a reward signal for alignment.
• Formalizes the 'ripple effect prediction' task and provides a standardized benchmark for evaluation.
• Demonstrates superior performance in explaining asset pricing residuals compared to RAG, zero-shot, and ICL baselines.
• Shows practical utility in portfolio management, achieving higher Sharpe ratios and lower drawdowns than traditional benchmarks.

Method

• Constructs time-varying knowledge graphs from unstructured financial data.
• Injects KG information into LLMs using lightweight adapter modules via instruction tuning.
• Uses Proximal Policy Optimization (PPO) to fine-tune the LLM backbone.
• Designs a reward function based on the correlation between predicted shocks and CAPM residuals.
• Evaluates explanatory power using regression analysis on asset pricing residuals.
• Compares performance against RAG, zero-shot, and in-context learning baselines.
• Implements a long-short portfolio strategy based on predicted event impacts.

Results & findings

• FinRipple achieves R2 up to 0.59 on Fama-French 5-factor residuals, significantly outperforming baselines (R2 < 0.25).
• The model shows statistical significance (p < 0.001) in explaining residual variance across multiple LLM architectures.
• Refusal-to-answer rates are significantly lower for FinRipple compared to zero-shot and ICL methods.
• Portfolio strategy using FinRipple predictions yields a Sharpe ratio of 1.153, outperforming Equal Weighting (0.882) and Markowitz (0.954).
• Maximum drawdown is reduced to -0.283, compared to -0.401 for Min-Variance Weighting.
• Ablation studies confirm the importance of supply chain and shared leadership relationships in KG construction.
• Parameter-efficient adaptation (adapters) allows temporal adaptation without retraining the full LLM backbone.

Data & experiments

• Dataset: 110,000 news articles for training and 10,000 for testing (S&P 500 companies, Jan 2020-Jun 2022).
• Knowledge Graphs: Constructed from leadership overlaps, mutual fund holdings, patent relationships, and supply chains.
• Baselines: RAG, Zero-Shot, In-Context Learning (ICL), and FinRipple without market alignment.
• Metrics: Explanatory power (R2) on CAPM and Fama-French residuals, ANOVA F-statistics, refusal-to-answer rates.
• Portfolio Backtest: Intraday strategy selecting top/bottom 10% predicted impact stocks, evaluated on daily return, Sharpe ratio, and max drawdown.

Limitations & caveats

• High computational cost: Training required over 9000 GPU hours.
• Dependency on KG quality: Performance relies on the accuracy and completeness of constructed knowledge graphs.
• Limited to S&P 500: Experiments focus on large-cap US stocks, potentially limiting generalizability.
• Static adapter reinitialization: Adapters are reinitialized for each fine-tuning step, which may impact temporal consistency.
• Refusal-to-answer rates: While improved, some models still exhibit refusal behavior in complex scenarios.

Curator notes

Discovery status: specific_domain_round2_promoted. Source links: FinKario: Event-Enhanced Automated Construction of Financial Knowledge Graph