Can We Estimate Novel Financial Risk Using LLM-Derived Risk Scores and LSTM Models?
Sudip Gupta and Hanboya Yan introduce an innovative framework that combines natural language processing (NLP) and deep learning to improve the prediction of firm-specific (idiosyncratic) volatility.
Gupta & Yan recognize a critical shortcoming in traditional financial risk models: the inability to effectively capture emerging “novel risks” such as ESG concerns, geopolitical tensions, and supply chain disruptions, all of which have grown in significance in today’s interconnected global economy.
Key Contributions and Strengths
The paper’s main contribution lies in its ability to extract and quantify qualitative risk narratives from corporate disclosures using large language models (LLMs). By translating subjective narratives into measurable risk scores, the authors create a new dimension of input that complements conventional financial indicators such as leverage, beta, and short interest.
These risk scores are then integrated into a long short-term memory (LSTM) neural network, a powerful tool for modeling time series data with temporal dependencies. The empirical analysis, spanning over 18,000 regulatory filings from 2015 to 2024, demonstrates that the integrated model achieves superior forecasting performance, with notable improvements in R² values and reduced mean squared error (MSE) compared to traditional models.
An important finding is that feature importance analyses confirm that these qualitative, nontraditional risk factors play a pivotal role in explaining firm-level volatility. This suggests that emerging risks—often dismissed as hard to quantify—hold significant predictive power and should not be ignored by investors or risk managers.
Practical Implications
This study has strong relevance for practitioners in risk management, portfolio management, and regulatory compliance. The ability to systematically identify and quantify qualitative risks provides a toolset for firms to enhance risk monitoring frameworks and proactively adjust strategies in response to evolving risk environments.
Moreover, this methodology reflects a broader trend in finance: the integration of unstructured data sources with machine learning techniques to extract deeper insights that are unavailable through traditional quantitative metrics alone.
Limitations and Areas for Future Research
The paper does acknowledge key limitations, including challenges inherent in large language models such as token limits, the need for more granular data, and the complexity of mapping narrative text to quantitative signals. Future research could build on this work by exploring additional forms of unstructured data, expanding the framework for high-frequency or real-time analysis, and testing across different sectors or under varying market conditions to assess robustness and generalizability.
Overall Assessment
This paper represents a meaningful advance in the evolving intersection of AI, NLP, and financial risk modeling. Its methodological rigor, empirical validation, and practical relevance make it an important contribution to the literature on volatility forecasting and risk management.
The authors convincingly demonstrate that emerging qualitative risks matter—and can be systematically incorporated into quantitative frameworks to improve prediction and understanding of market volatility. As firms increasingly confront ESG pressures, geopolitical uncertainty, and supply chain fragility, tools like the one proposed in this study will likely become essential for forward-looking risk managers and sophisticated investors.
In short, this research highlights how the fusion of cutting-edge AI with traditional financial modeling offers a more comprehensive and nuanced approach to measuring and managing risk in a rapidly changing world.
Download the full paper: Using Large Language Models to Estimate Novel Risk: Impact on Volatility | Portfolio Management Research
