Finance and Large Language Models

Inhaltsverzeichnis

1. Frontmatter

2. Large Language Models in Finance: An Overview

   Paul Moon Sub Choi, Seth H. Huang, Qishu Wang

   Abstract

   The rapid advancement of large language models (LLMs) is revolutionizing industries, with finance emerging as one of the most promising beneficiaries. Financial LLMs (FinLLMs), developed on their foundations, can leverage advanced natural language processing techniques to process and generate insights from vast volumes of unstructured financial data. Particularly in specialized areas like quantitative investing, FinLLMs are poised to redefine the landscape by emulating the decision-making process of top traders. They can more efficiently capture market expectations, evaluate the impacts of market events, and aid investors across a wide range of investment practices. However, systematic research in the field of FinLLMs remains in its early stages. This paper provides a comprehensive overview of FinLLMs, aiming to encourage broader exploration of their mature applications in finance. The main content is summarized as follows: Firstly, we present a chronological overview tracing the evolution from general-domain pre-trained language models (PLMs) to specialized FinLLMs, highlighting critical advancements such as FinBERT, BloombergGPT, and FinMA. Secondly, we compare major FinLLMs by examining their training methods, datasets, and corresponding fine-tuning strategies. Thirdly, we summarize the characteristics and performance evaluations of seven benchmark financial NLP tasks. In addition, we explore the practical applications of FinLLMs in traditional finance and behavioral finance. Finally, we discuss the challenges and opportunities in the adoption of FinLLMs, including issues like data privacy and ethical considerations, while proposing directions for future research.

3. Housing Price Estimation and Reasoning Based on a Large Language Model

   Seongeun Bae, Leehyun Jung, Sukyung Nam, Sihyun An, Kwangwon Ahn

   Abstract

   This study investigates the applicability of a large language model (LLM) to housing price appraisal in terms of predictive power and explainability. We first transform a hedonic dataset into a convertible format to construct the LLM-based appraisal framework using the established prompt engineering. We then compare the results to those obtained using a traditional hedonic pricing model. Our findings reveal that LLM outperforms the traditional benchmark model concerning two accuracy measures (i.e., root mean square error and R² value) in appraising housing prices. This outcome indicates the substantial capability of LLM for seizing nonlinearity in the hedonic dataset. Furthermore, the LLM-based appraisal framework provides three-dimensional interpretations, including (1) the directional impacts, (2) the qualitative importance of the hedonic variables concerning housing prices, and (3) narrative reasoning for the appraised prices. These findings reinforce that the proposed LLM-based valuation model is a potential tool for understanding the mechanism of housing prices. Investors can implement our framework to estimate properties and support decision-making through explainable LLM results. Moreover, policymakers can benchmark our results when developing monitoring systems and designing transparent real estate markets.

4. Advancing Quantitative Trading Strategies Using Fine-Tuned Open-Source Large Language Models: A Hybrid Approach with Numerical and Textual Data Integration Using RAG and LoRA Techniques

   Seth H. Huang, Jimin Kim, Ka Lok Kellogg Wong

   Abstract

   This paper explores the latest methodologies for fine-tuning open-source Large Language Models (LLMs) in enhancing quantitative trading strategies by integrating numerical data (e.g., historical prices, technical indicators) with textual data (e.g., news, earnings reports, social media sentiment). We employ Retrieval-Augmented Generation (RAG) with a vector database to efficiently handle and contextualize textual data, alongside Low-Rank Adaptation (LoRA) techniques for cost-effective and scalable model fine-tuning. The proposed approach aims to create a hybrid trading model that combines the predictive power of LLMs with traditional quantitative methods, improving accuracy and adaptability in financial markets. This study details the implementation process, highlighting practical innovations such as the integration of real-time data pipelines and adaptive model tuning. Experimental results show significant improvements in predictive accuracy and risk-adjusted returns, demonstrating the practical value of these advanced fine-tuning methodologies in finance.

5. Foundations of LLMs and Financial Applications

   Yoonseo Chung, Jeonghyun Kim, MiYeon Kim, Minsuh Joo, Hyunsoo Cho

   Abstract

   The integration of Large Language Models (LLMs) into the financial industry represents a transformative advancement in artificial intelligence, addressing the complexities of data-driven finance. This chapter explores how cutting-edge LLMs can be aligned with financial practices to enhance efficiency and foster innovation in financial services. The discussion begins with an overview of LLM, including their architecture, training processes, and the datasets they leverage. It then examines finance-specific adaptations, such as FinBERT and BloombergGPT, which are tailored to address domain-specific challenges. The chapter also addresses key challenges in applying LLMs to the financial domain, such as real-time data integration, and evaluates potential solutions, including retrieval-augmented generation (RAG). By analyzing these innovations and challenges, the chapter envisions a future where LLMs redefine the landscape of financial technology.

6. Voluntary Sustainability Disclosure and Third-Party Assurance: A Large Language Model Perspective

   SoHyeon Kang, Sewon Kwon

   Abstract

   This chapter explores the influence of third-party assurance on the sentiment and subjectivity of corporate sustainability reports. As sustainability reporting grows in importance for corporate transparency and stakeholder engagement, the role of assurance—whether limited or reasonable and provided by audit or non-audit firms—has become increasingly critical in shaping report narratives. Utilizing advanced sentiment analysis methodologies, including BERT-based models, we analyze the tonal qualities and factual content of sustainability reports. Findings indicate that reports without assurance exhibit more positive sentiment, while those with limited or reasonable assurance reflect varying degrees of narrative objectivity and sentiment neutrality. Reports assured by audit firms tend to convey more neutral and comprehensive narratives, emphasizing factuality over subjective tones, compared to those assured by non-audit entities. This analysis contributes to the understanding of how assurance practices impact the perceived credibility and faithful representation of narrative sustainability disclosures. By combining natural language processing insights with empirical data, our study underscores the transformative role of assurance in enhancing nonfinancial disclosures quality and fostering accountability. The results provide policy and practical implications for the discussion on mandating third-party assurance of nonfinancial disclosures.

7. Verbal Femininity and CEOs Compensation

   Sang-Joon Kim, Juil Lee

   Abstract

   This study investigates how gender biases among board directors influence CEO compensation, focusing on the role of male CEOs’ verbal femininity. Given that verbal expression is another essential way to reveal feminine characters of CEOs, the femininity saliently cued in CEOs’ verbalizing habits can bring more distorted perception, which leads to the boards’ evaluation on their contributions. This is because verbal femininity, characterized by nurturing, empathetic, benevolent, and collaborative communication styles, is argued to conflict with traditional masculine stereotypes of leadership. Drawing on role congruity theory, we contend that CEOs who exhibit verbal femininity are undervalued in compensation decisions due to biases against traits perceived as incongruent with the prototype of effective leadership. In this study, we specify the feminine characteristics which induce decision errors by using a Large Language Model (LLM). Acknowledging that LLMs are subject to pre-existing social prejudices and gender biases, we utilize this attribute of the models to develop a way to capture gender biases (especially femininity). In this study, we employ an LLM-driven algorithm to measure the extent to which a certain CEO’s words in official settings show femininity. Using this measure, we examine how the CEO’s verbal femininity can affect the assessment of CEO quality in the boardroom, determining the level of CEO compensation. Our findings contribute to the broader discourse on implicit gender biases in determining CEO compensation.

8. Integrating LLM-Based Time Series and Regime Detection with RAG for Adaptive Trading Strategies and Portfolio Management

   Chenkai Li, Chi Ho Roger Chan, Seth H. Huang, Paul Moon Sub Choi

   Abstract

   This paper explores the latest methodologies for fine-tuning open-source Large Language Models (LLMs) to enhance quantitative trading strategies by integrating numerical data (e.g., historical prices, technical indicators) with textual data (e.g., news, earnings reports, social media sentiment). We employ Retrieval-Augmented Generation (RAG) with a vector database to efficiently handle and contextualize textual data, enabling LLMs to derive actionable insights from both structured and unstructured data. The proposed approach focuses on fully fine-tuning smaller models, such as GPT-4o Mini, for cost-effective and scalable applications in finance. The study aims to create a hybrid trading model that combines the predictive power of LLMs with traditional quantitative methods, improving accuracy and adaptability in financial markets. Key innovations include the integration of real-time data pipelines and adaptive model tuning. Experimental results demonstrate significant improvements in predictive accuracy and risk-adjusted returns, showcasing the practical value of these advanced fine-tuning methodologies in finance.

9. Empirical Factor Identification for Artificial Intelligence in Finance: Indian Evidence

   Rohit Kaushik

   Abstract

   In the fast-changing landscape of financial markets, artificial intelligence is making inroads into it. Artificial intelligence is a concept or technological advancement which aims to mirror intelligence exhibited by the human beings and executes all the jobs to the perfection. Financial or investment domain is not an exception to this. Newer technologies are making their way to the horizon like Large Language Models (LLMs) as they present themselves into innovative and interactive methods of extracting information for making informed investment decisions. Before making any investment decision one has to look at various aspects and pay attention to various reports and constantly monitor the changes in stock prices, which is a very difficult job every person can’t claim to master this art, in this artificial intelligence comes to the rescue of such persons who can’t perform such tasks carefully. The present study is an attempt in this regard and it consists of respondents who are living in Delhi national capital of India. Data was collected with the help of questionnaire and analyzed by employing multiple regression method to ascertain the effect of artificial intelligence on investment decisions in addition to overconfidence bias which was another factor studied with artificial intelligence.

10. Federated and Decentralized Finance: Decentralized Reward Mechanisms for Advanced AI Learning

    Hyoseok Jang, Sangchul Lee, Haneol Cho, Chansoo Kim

    Abstract

    Federated Learning (FL) is a distributed machine learning paradigm in which multiple clients collaboratively train a global model while keeping their private data local [2, 6, 8].

11. AI-Driven Financial Chart Analysis with Benchmarks: A Domain-Specific Large Language Model Approach

    Hyoseok Jang, Sangchul Lee, Haneol Cho, Chansoo Kim

    Abstract

    Since the release of GPT-3.5, Large Language Models (LLMs) have taken a dramatic leap forward, surpassing previous state-of-the-art achievements in tasks such as reading comprehension, code generation, and creative text composition [1, 2].