R2D-LLM: Retrieval-Augmented Large Language Model Driven Dynamic Risk Diffusion Network for Private Enterprise Distress Prediction and Risk-Aware Investment Decision Support

Authors

  • Wenxuan Liu Computer Science, University of California, Los Angeles,Los Angeles,CA,USA
  • Chongli Wu Master in Applied Finance, Pepperdine University, Malibu, CA, USA
  • Yuchen Gan Master of Science, Columbia University, New York, NY, USA

DOI:

https://doi.org/10.66238/fsrma143

Keywords:

Large Language Model, Retrieval-Augmented Generation, Dynamic Risk Diffusion, Private Enterprise Distress Prediction, Risk-Aware Decision Support, Financial Artificial Intelligence

Abstract

Private enterprises in venture capital (VC) and private credit markets are difficult to assess because limited disclosure, incomplete credit histories, and rapidly changing operating conditions restrict the usefulness of conventional ratio-based models. Important warning signals are often dispersed across financing announcements, regulatory disclosures, business reports, and market news, while enterprise distress typically develops through interacting events rather than a single observation. This study proposes R2D-LLM, a retrieval-augmented large language model driven dynamic risk diffusion framework for private enterprise distress prediction and risk-aware investment decision support. The framework uses a RAG-enhanced LLM to convert heterogeneous documents into structured risk events containing category, severity, timestamp, semantic description, and confidence information. A dynamic risk event graph and Risk Diffusion Transformer then model temporal dependencies, risk accumulation, cross-category propagation, and attenuation. Lifecycle embeddings and enterprise-level financial and contextual features are integrated to estimate twelve-month distress risk across different financing stages. On the reported multi-source dataset containing 12,846 enterprises, 186,520 documents, and 94,317 extracted risk events, R2D-LLM achieves 92.8% accuracy, a 90.6% F1-score, and an AUC of 0.936, outperforming the evaluated machine-learning, deep-learning, Transformer, and financial-LLM baselines. Reported ablation results suggest that retrieval-enhanced event extraction and dynamic diffusion modeling provide the largest contributions. These findings support R2D-LLM as a dynamic enterprise risk-assessment framework and provide a basis for subsequent validation of downstream investment strategies.

References

[1] Brown, Tom, et al. "Language models are few-shot learners." Advances in neural information processing systems 33 (2020): 1877-1901.

[2] Yang, Hongyang, Xiao-Yang Liu, and Christina Dan Wang. "Fingpt: Open-source financial large language models." arXiv preprint arXiv:2306.06031 (2023).

[3] Wu, Shijie, et al. "Bloomberggpt: A large language model for finance." arXiv preprint arXiv:2303.17564 (2023).

[4] Lewis, Patrick, et al. "Retrieval-augmented generation for knowledge-intensive nlp tasks." Advances in neural information processing systems 33 (2020): 9459-9474.

[5] Gao, Yunfan, et al. "Retrieval-augmented generation for large language models: A survey." arXiv preprint arXiv:2312.10997 (2023).

[6] Wang, Tangtang, Kaijie Zhang, and Kuangcong Liu. "A knowledge graph and deep learning-based semantic recommendation database system for advertisement retrieval and personalization." arXiv preprint arXiv:2601.00833 (2025).

[7] Hawkes, Alan G. "Spectra of some self-exciting and mutually exciting point processes." Biometrika 58.1 (1971): 83-90.

[8] Mei, Hongyuan, and Jason M. Eisner. "The neural hawkes process: A neurally self-modulating multivariate point process." Advances in neural information processing systems 30 (2017).

[9] Zuo, Simiao, et al. "Transformer hawkes process." International conference on machine learning. PMLR, 2020.

[10] Yoon, Kanghoon, et al. "Learning multivariate hawkes process via graph recurrent neural network." Proceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining. 2023.

[11] Mestiri, Sami. "Credit scoring using machine learning and deep Learning-Based models." Data Science in Finance and Economics 4.2 (2024): 236.

[12] Mojdehi, Kosar Farajpour, Babak Amiri, and Amirali Haddadi. "A novel hybrid model for credit risk assessment of supply chain finance based on topological data analysis and graph neural network." IEEE Access 13 (2025): 13101-13127.

[13] Jiang, Gaozhe, et al. "Investment advisory robotics 2.0: Leveraging deep neural networks for personalized financial guidance." 2025 International Joint Conference on Neural Networks (IJCNN). IEEE, 2025.

Downloads

Published

2026-10-31