Earticle

초록

영어

Hallucinations in Large Language Model (LLM)-generated financial summaries pose substantial risks to algorithmic trading systems, yet most existing studies feed unverified model outputs directly into trading pipelines. This paper proposes an end-to-end framework that converts unstructured financial news into reliable quantitative trading signals while systematically controlling hallucination risk. The framework combines a finance-specific Chain-of-Density (CoD) summarization agent with a three-layer hallucination verification pipeline comprising TF-IDF-based evidence retrieval, deterministic numeric validation, and DeBERTa-v3-large NLI-based semantic consistency verification. Verified signals are subsequently processed by a block-memory trading agent with configurable rebalancing windows. Ablation experiments across MSFT, BA, and DIS spanning 2019–2021 demonstrate that full verification (Exp3) consistently outperforms unverified baselines across cumulative return, Sharpe ratio, Sortino ratio, and MDD. The 20-day memory window yields optimal risk-adjusted performance across all tickers, confirming that hallucination control and memory configuration jointly determine trading pipeline reliability.

목차

Abstract
1. Introduction
2. Related Works
2.1 Traditional Financial Sentiment Analysis
2.2 Advances in Deep Learning-Based Financial News Processing
2.3 Technical Evolution of Finance-Specific Pretrained Models
2.4 LLM-Based Autonomous Trading and Hallucination Risk
3. Preliminaries
3.1 Chain of Density Prompting
3.2 TF-IDF-Based Document Retrieval
3.3 DeBERTa-v3-Based Natural Language Inference
3.4 Chain-of-Thought Prompting
4. Proposed Method
4.1 News Collection & Preprocessing
4.2 Finance-Specific CoD Agent
4.3 Hallucination Verification Framework
4.4 Hierarchical Structured Signal Extraction Agent
4.5 Block-Memory Trading Agent
5. Experiments
5.1 Dataset
5.2 Experimental Setup
5.3 Evaluation Metrics
6. Results
6.1 Memory Window Comparison Results
6.2 Ablation Study Results of the Hallucination Verification Framework
7. Conclusion
References

키워드

저자

• Hunmin Jung [ Tech University of Korea Department of IT Management ]
• Seungwon Park [ Tech University of Korea Department of IT Management ]
• Minsu Kang [ Tech University of Korea Department of Business Administration ]
• Junhyoek Jang [ Tech University of Korea Department of Data Science and Business Administration ]
• Chaehoon Yun [ Tech University of Korea Department of Business Admin ]
• Yuhee Kwon [ Tech University of Korea Department of Business Administration ] Corresponding Author

참고문헌

발행기관

간행물

표지보기 관심저널 등록

• 간행물명

  한국경영정보학회 정기 학술대회 [KMIS Conference]

• 간기

  반년간

• 수록기간

  1990~2026

• 십진분류

  KDC 325 DDC 658