Agent AI for Finance From Financial Argument Mining to Agent-Based Modeling (Chung-Chi Chen, Hiroya Takamura) (Z-Library)

📄 File Format: PDF

💾 File Size: 1.3 MB

📖 Read Online ⬇️ Download

Registered users can read the full content for free

Register as a Gaohf Library member to read the complete e-book online for free and enjoy a better reading experience.

📄 Page 1

SpringerBriefs in Intelligent Systems Artificial Intelligence, Multiagent Systems, and Cognitive Robotics Chen · Takam ura A gent A I for Finance Chung-Chi Chen · Hiroya Takamura Agent AI for Finance From Financial Argument Mining to Agent-Based Modeling

📄 Page 2

SpringerBriefs in Intelligent Systems Artificial Intelligence, Multiagent Systems, and Cognitive Robotics Series Editors Gerhard Weiss, Maastricht University, Maastricht, The Netherlands Karl Tuyls, University of Liverpool, Liverpool, UK; Google DeepMind London, UK Editorial Board Felix Brandt, Technische Universität München, Munich, Germany Wolfram Burgard, Albert-Ludwigs-Universität Freiburg, Freiburg, Germany Marco Dorigo , Université Libre de Bruxelles, Brussels, Belgium Peter Flach, University of Bristol, Bristol, UK Brian Gerkey, Open Source Robotics Foundation, Mountain View, USA Nicholas R. Jennings, Imperial College London, London, UK Michael Luck, King’s College London, London, UK Simon Parsons, City University of New York, New York, USA Henri Prade, IRIT, Toulouse, France Jeffrey S. Rosenschein, Hebrew University of Jerusalem, Jerusalem, Israel Francesca Rossi, University of Padova, Padua, Italy Carles Sierra, IIIA-CSIC Cerdanyola, Barcelona, Spain Milind Tambe, University of Southern California, Los Angeles, USA Makoto Yokoo, Kyushu University, Fukuoka, Japan

📄 Page 3

This series covers the entire research and application spectrum of intelligent systems, including artificial intelligence, multiagent systems, and cognitive robotics. Typical texts for publication in the series include, but are not limited to, state-of-the- art reviews, tutorials, summaries, introductions, surveys, and in-depth case and application studies of established or emerging fields and topics in the realm of compu- tational intelligent systems. Essays exploring philosophical and societal issues raised by intelligent systems are also very welcome.

📄 Page 4

Chung-Chi Chen · Hiroya Takamura Agent AI for Finance From Financial Argument Mining to Agent-Based Modeling

📄 Page 5

Chung-Chi Chen Artificial Intelligence Research Center AIST Koto-ku, Tokyo, Japan Hiroya Takamura Artificial Intelligence Research Center AIST Koto-ku, Tokyo, Japan ISSN 2196-548X ISSN 2196-5498 (electronic) SpringerBriefs in Intelligent Systems ISBN 978-3-031-94686-8 ISBN 978-3-031-94687-5 (eBook) https://doi.org/10.1007/978-3-031-94687-5 This work was supported by National Institute of Advanced Industrial Science and Technology. © The Editor(s) (if applicable) and The Author(s) 2025. This book is an open access publication. Open Access This book is licensed under the terms of the Creative Commons Attribution 4.0 International License (http://creativecommons.org/licenses/by/4.0/), which permits use, sharing, adaptation, distribu- tion and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license and indicate if changes were made. The images or other third party material in this book are included in the book’s Creative Commons license, unless indicated otherwise in a credit line to the material. If material is not included in the book’s Creative Commons license and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. The use of general descriptive names, registered names, trademarks, service marks, etc. in this publication does not imply, even in the absence of a specific statement, that such names are exempt from the relevant protective laws and regulations and therefore free for general use. The publisher, the authors and the editors are safe to assume that the advice and information in this book are believed to be true and accurate at the date of publication. Neither the publisher nor the authors or the editors give a warranty, expressed or implied, with respect to the material contained herein or for any errors or omissions that may have been made. The publisher remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. This Springer imprint is published by the registered company Springer Nature Switzerland AG The registered company address is: Gewerbestrasse 11, 6330 Cham, Switzerland If disposing of this product, please recycle the paper.

📄 Page 6

This book is dedicated to all contributors in this field.

📄 Page 7

Preface Financial documents contain numerous causal inferences and subjective opinions. In our previous book, “From Opinion Mining to Financial Argument Mining,”1 we discussed understanding financial documents in a fine-grained manner, particularly those containing opinions. We highlighted several future directions, such as financial argument mining, multimodal opinion understanding, and analysis generation. We anticipated a lengthy journey for these topics. However, since 2022, ChatGPT and large language models (LLMs) have shown promising advancements, motivating us to write the second book that falls under the financial NLP topic. This book provides an overview of the current state of financial argument mining and financial text generation, and presents our thoughts on the blueprint for NLP in finance in the Agent AI era. Agent-based AI systems have been widely discussed since the advent of LLMs. This book aims to equip researchers and practitioners with the latest methodologies, concepts, and frameworks for developing, deploying, and evaluating AI agents with capabilities in multimodal understanding, decision-making, and interaction. It places a special emphasis on human-centered decision-making and multi-agent cooperation in financial applications. We survey the current landscape and discuss future research and development directions. Targeting a wide audience, from students to seasoned researchers in AI and finance, this book offers an overview of recent trends in Agent AI for finance. It provides a foundation for students to understand the field and design their research direction while inviting experienced researchers to engage in discussions on open research questions informed by pilot experimental results. 1 Open Access: https://link.springer.com/book/10.1007/978-981-16-2881-8. vii

📄 Page 8

viii Preface Although this book focuses on financial applications, the discussed concepts and methods can also be applied to other real-world applications by integrating domain- specific characteristics. We look forward to seeing new findings and more novel extensions based on the proposed ideas.2 Koto-ku, Japan March 2025 Chung-Chi Chen Hiroya Takamura 2 We presented a tutorial at ECAI-2024 based on the content of this book. The slides are available at: https://sites.google.com/view/finagent/home.

📄 Page 9

Acknowledgments Many researchers have assisted us in writing this book. We would like to thank all the students and collaborators of the NLPFin—Prof. Hsin- Hsi Chen (National Taiwan University), Prof. Chenghua Lin (University of Manch- ester), Prof. Kiyoshi Izumi (University of Tokyo), Dr. Hen-Hsen Huang (Academia Sinica), Prof. Noriko Kando (National Institute of Informatics), Prof. Sudip Kumar Naskar (Jadavpur University), Prof. Yusuke Miyao (University of Tokyo), Prof. Ichiro Kobayashi (Ochanomizu University), Prof. Ryutaro Ichise (Tokyo Insti- tute of Technology), Dr. Natthawut Kertkeidkachorn (Japan Advanced Institute of Science and Technology), Dr. Rungsiman Nararatwong (AIST), Dr. Ramon Ruiz Dolz (Universitat Politècnica de València), Jui Chu (National Taiwan University), Pei-Wei Kao (National Taiwan University), Tsun-Hsien Tang (National Taiwan University), JianTao Huang (National Taiwan University), Ting-Wei Hsu (National Taiwan University), Yi-Ting Liu (National Taiwan University), Ming-Xuan Shi (National Taiwan University), Chr-Jr Chiu (National Taiwan University), Wei-Lin Chen (National Taiwan University—AIST Intern), Sin-Han Yang (National Taiwan University), Tomas Goldsack (University of Sheffield—AIST Intern), Xingwei Qu (University of Manchester), Yuyang Cheng (University of Manchester), Yi-Ning Juan (National Taiwan University), Yu-Min Tseng (National Taiwan University), Chin-Yi Lin (National Taiwan University), Tsung-Hsuan Pan (National Taiwan University), Takehiro Takayanagi (The University of Tokyo—AIST RA), Hsiu-Hung Lee (National Yang Ming Chiao Tung University), Bo-Wei Chen (National Yang Ming Chiao Tung University), Hanwool Lee (NCSOFT), Yung-Yu Shih (National Taiwan University), Sohom Ghosh (Jadavpur University), Hsiu-Hsuan Yeh (National Taiwan University), Cheng-KuangWu (National Taiwan University), Prof. Yohei Seki (University of Tsukuba), Dr. Juyeon Kang (3DS Outscale), Anaïs Lhuissier (3DS Outscale), Prof. Min-Yuh Day (National Taipei University), and Prof. Yu-Lieh Huang (National Tsing Hua University). On the funding side, this book was partially supported by a project JPNP20006, commissioned by the New Energy and Industrial Technology Development Organi- zation (NEDO) and JSPS KAKENHI Grant Number 23K16956. March 2025 Chung-Chi Chen Hiroya Takamura Competing Interests The authors have no competing interests to declare that are relevant to the content of this manuscript. ix

📄 Page 10

Contents 1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1 1.1 From Opinion Mining to Financial Argument Mining . . . . . . . . . . . . . 1 1.2 From Financial Argument Mining to Agent-Based Modeling . . . . . . 2 1.3 Why Study Agent AI? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3 1.4 Overview of the Book . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7 References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7 2 Financial Argument Mining . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9 2.1 Argument Structure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9 2.2 Forward-Looking Argument Mining . . . . . . . . . . . . . . . . . . . . . . . . . . . 12 2.3 Argument Quality and Forecasting Skill Assessment . . . . . . . . . . . . . . 15 2.4 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17 References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18 3 Single-Agent/Model Design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21 3.1 Learning from Human Insights . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21 3.2 Retrieval-Augmented Generation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24 3.3 Model Editing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27 3.4 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29 References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30 4 Multi-agent Interaction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35 4.1 Multi-round Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35 4.2 Hierarchical Decision-Making . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38 4.3 Human Behavior Simulacra . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40 4.4 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43 References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44 xi

📄 Page 11

xii Contents 5 Multi-scale Model Synergy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47 5.1 Data Augmentation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47 5.2 Dynamic Interaction Loop . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50 5.3 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53 References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56 6 Generative AI Application Scenarios . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59 6.1 Extension of Impact Duration Inference . . . . . . . . . . . . . . . . . . . . . . . . 59 6.2 Opinion Ranking . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62 6.3 Numeracy and Reasoning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64 6.4 Creative Agent . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66 6.5 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67 References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68 7 Looking to the Future . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71 7.1 Progress on Previously Proposed Research Directions . . . . . . . . . . . . 71 7.2 Future Directions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75 7.3 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80 References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82

📄 Page 12

Chapter 1 Introduction Intelligence encompasses understanding, reasoning, planning, inference, decision- making, and more. In our previous book [ 5], we focused on understanding financial documents, particularly those containing opinions, from an information extraction perspective. In this book, we extend our discussions to include argument-mining notions and further explore reasoning, planning, inference, and decision-making. This chapter provides an overview of the book. In Sect. 1.1, we recap the discussions from our previous book [ 5] and highlight the argument mining concept that will be discussed in this book. Section 1.2 outlines the roadmap from understanding human- written content to simulating human discussions. In Sect. 1.3, we share our thoughts on why Agent AI is a promising research topic for financial applications. Finally, we provide an overall structure of this book in Sect. 1.4. 1.1 From Opinion Mining to Financial Argument Mining In our previous book [ 5], we proposed understanding financial opinions through the following twelve components: target entity, market sentiment, opinion holder, publishing time, validity period of an opinion, market information set, analysis aspect, degree of market sentiment, a set of claims, a set of premises, opinion quality, and influence power. Although we provided some results and promising directions at that time, several components remained unexplored, particularly the validity period, argument-based analysis, and opinion quality. In this book, we aim to fill this gap and share some experimental results. We will primarily use argument-mining notions for these discussions; thus, we first recap the idea of financial argument mining. In contrast to opinion mining and sentiment analysis, which mainly focus on clas- sifying a given statement as positive/negative (bullish/bearish), argument mining provides a more in-depth understanding of an opinion. For example, the causal rela- tionship and the logic between premises and claims can provide clues for estimating © The Author(s) 2025 C.-C. Chen and H. Takamura, Agent AI for Finance, SpringerBriefs in Intelligent Systems, https://doi.org/10.1007/978-3-031-94687-5_1 1

📄 Page 13

2 1 Introduction the strength of an argument. Based on Toulmin’s argumentative model [ 20], we can separate a narrative into two basic units, claim and premise, where the claim is the subjective view of the investor, and the premise is the objective act used to support the claim. With this idea, we can transform a financial analysis, such as a professional research report for a particular company, into directed graphs called argumentation structures. The argumentation structure has been widely used in other domains, such as debate quality assessment [ 15] and persuasive essay evaluation [ 21], but has yet to be discussed in financial analysis. In Chap. 2, we provide in-depth discussions on how to use argument-based analysis for financial purposes. A major difference between financial arguments and arguments in other domains is that people always discuss the future in financial analysis. That is, we encounter many forward-looking statements in financial narratives. For example, investors discuss the company’s future operations and possible stock price movements daily. This leads to two topics: the duration for which a given premise will influence companies’ operations and the validity period of a given claim. For example, an event in 2020 may have little influence in 2025, and investors may not consider an opinion from 2019 when making decisions in 2026. Following this line of thought, we also propose adding a temporal dimension to the argument-based financial analysis in Chap. 2. We believe that decomposing an analysis from the argument-mining perspective can lead to more insights for both human beings and machine learning models, and we hope our discussion will inspire more researchers to explore the potential of financial argument mining. 1.2 From Financial Argument Mining to Agent-Based Modeling Simulating human behaviors and actions is one of the major goals of AI systems. After achieving the capability of understanding financial documents, generating (rea- soning, planning, and inference), and making decisions becomes the next challenge. In 2021, it was still difficult for models to generate fluent financial analysis reports. Thanks to the development of generative models, we soon obtained superior LLMs for solving the fluency problem in 2022. In the past two years, researchers have started to discuss the potential of LLM agents, and our goal of generating analysis reports in our previous book [ 5] has become possible. Therefore, in Chaps. 3 and 4, we focus on how to use an agent or agents to simulate the behaviors of professionals in the financial sector, especially those writing analysis reports and making trading decisions. The definition of agent-based AI systems (Agent AI) is still open to discussion. It has been defined as “a class of interactive systems that can perceive visual stim- uli, language inputs, and other environmentally grounded data, and can produce meaningful embodied actions” [ 10]. In this book, we define Agent AI for finance as a system involving the interaction among multiple agents/models that can accept

📄 Page 14

1.3 Why Study Agent AI? 3 multimodal inputs (video, audio, image, and text) and is able to output useful infor- mation for the human-centered decision-making process in financial contexts. For example, when generating an analysis report based on the audio and transcript of an earnings conference call, several LLMs would be asked to examine different aspects, and then another LLM would be asked to summarize and make inferences based on the feedback from different LLMs. We discuss this kind of Agent AI in Chap. 4. The definition of agent-based AI systems (Agent AI) is still open to discussion. It was defined as “a class of interactive systems that can perceive visual stimuli, lan- guage inputs, and other environmentally grounded data, and can produce meaningful embodied actions” [ 10]. In this book, we define Agent AI for finance as a system involving the interaction among multiple agents/models that can accept multimodal inputs (video, audio, image, and text) and is able to output useful information for the human-centered decision-making process. For example, when generating an analy- sis report based on the audio and transcript of an earnings conference call, several LLMs would be asked to examine different aspects, and then another LLM would be asked to summarize and make inferences based on the feedback from different LLMs. This is a kind of agent AI we will discuss in Chap. 4. Agent-based modeling, which aims to simulate interactions in the real world, is a long-term topic of discussion in economics and finance [ 1]. Traditionally, agent- based modeling has always been done using several equations with different hyper- parameter settings. With LLMs, we are exploring whether it is possible to conduct simulations in a more natural way. That is, we plan to let LLMs play different roles and interact with other LLMs using natural language. For example, we can simulate people’s reactions to a rate hike in real estate and the stock market. If an LLM can simulate the decision of a given role, the interaction among LLMs would be very close to real-world outcomes. The same ideas can be applied to several tasks that were previously discussed with numerical simulation, such as consumer behavior and election simulation. Using LLMs for agent-based modeling enables researchers to observe changes in the simulated society and also provides explainable tracing routes for the changes. Although this is still in an early stage, we provide some discussions in Chap. 4. 1.3 Why Study Agent AI? After explaining our journey and vision from opinion mining to agent-based modeling in finance, we now share our perspective on Agent AI and its promising future direc- tion. Figure 1.1 illustrates our thoughts on this topic. AI models have traditionally been employed as tools for automation. For example, we can train a supervised model for sentiment analysis. By analyzing numerous social media posts, the model can determine the sentiments of social media users, effectively summarizing social media opinions [ 2]. Additionally, a model can be trained with annotated data to extract key information from financial documents, enabling investors to quickly grasp important

📄 Page 15

4 1 Introduction Fig. 1.1 Our vision of Agent AI information [ 3, 4, 6, 7]. While these AI models may achieve high performance for specific tasks, they are generally limited to only one or a few tasks. Unlike traditional AI models, LLMs acquire general capabilities after training on large datasets. In other words, AI models are analogous to a student who has specifi- cally learned one chapter or skill, such as addition or a particular task. These models excel at solving problems within their specific domain because they have been trained solely for that purpose. For example, a model might be trained exclusively to perform addition, thereby performing exceptionally well when dealing with addition prob- lems. In contrast, LLMs resemble a student who has received a broad foundational education, encompassing not only addition but also other mathematical concepts, and perhaps even knowledge from other subjects. These models possess a wider range of abilities, enabling them to handle various types of problems, although they may not always achieve perfect accuracy in every case. They are capable of answering open- ended questions in a variety of contexts, similar to how a student might apply their knowledge in different situations. Following this line of thought, it is conceivable that LLMs can be trained for different fields, much like students choosing a major in college. Taking this a step further, intelligence encompasses not only natural language processing (NLP) but also the ability to plan, learn, make decisions, reason, and

📄 Page 16

1.3 Why Study Agent AI? 5 more. This leads to an extension of LLMs: the AI Agent. The concept of an AI agent is to empower LLMs to divide a task into several subtasks, iteratively check the generated text from different perspectives, select appropriate tools for solving problems, and so on. Unlike the one-step problem-solving approach of traditional AI models or LLMs, AI agents should simulate human behaviors in problem-solving, such as multi-step planning and reasoning. To implement a multi-step problem- solving framework, an admin agent is required to select AI agents to form a team and guide the discussions among them. The admin agent also communicates with users and may require personalization capabilities. This multi-step framework signifies a transition from the “AI as Tool” era to the “AI as Partner” era, as illustrated in Fig. 1.1. Specifically, when AI is seen as a tool, it is primarily used to accomplish specific tasks. In this mode, AI executes tasks assigned by humans but doesn’t operate beyond the given instructions. Humans rely on AI for automation, but ultimate decision- making and creativity remain in human control. For example, spell check or grammar correction in writing software can detect errors, but humans decide whether to accept the AI’s suggestions. AI in photo editing tools can automatically enhance photos based on patterns, but the user has the final say on adjustments and edits. However, when AI is seen as a partner, it plays a more proactive and collaborative role. AI not only executes tasks but also participates in the thinking and decision-making process. It helps solve creative and strategic problems, acting more as a co-worker than just a tool. For example, in healthcare, an AI partner provides diagnostic suggestions and treatment options based on patient history and the latest medical research, and doctors and AI collaborate to find the best treatment plan. In financial analysis, an AI partner doesn’t just analyze data but also offers insights into market trends, helps executives predict future outcomes, and guides decision-making collaboratively. In sum, “AI as a tool” is limited to executing tasks assigned by humans, focusing on specific operations without influencing decisions or creative processes. “AI as a partner” works alongside humans, engaging in decision-making, offering creative input, and contributing to strategic problem-solving in a collaborative manner. Recent studies have demonstrated the potential of the multi-agent framework. For example, instead of relying on a single LLM to generate code, employing mul- tiple agents to work collaboratively can yield better outcomes in software develop- ment [ 12]. These agents assume roles such as product managers, architects, project managers, and engineers and interact with one another to refine and generate code based on the given software requirements. Another significant application scenario is human behavior simulation [ 16]. This includes simulating the daily life of a small village or human behaviors in various professional contexts, such as buyer-seller interactions. By accurately capturing the persona of each market participant, we can approach a realistic market simulation. Furthermore, it is possible to observe whether AI agents exhibit behaviors similar to humans, such as overconfidence [ 19], overreaction [ 9], or herding behavior [ 8]. AI agents introduce a new dimension to agent-based modeling, as they can now communicate using natural language rather than the numerical methods previously employed. This advancement opens up oppor- tunities to revisit a wide range of agent-based modeling research by employing AI agents.

📄 Page 17

6 1 Introduction Fig. 1.2 Illustration of multi-scale model synergy In addition to discussions focused solely on LLM-based agent interaction [ 11], we propose the concept of multi-scale model synergy in Chap. 5. Our goal is to bridge the efforts on pre-trained language models (PLMs) and LLMs to develop an advanced multi-agent framework. As illustrated in Fig. 1.1, previous studies have primarily concentrated on either AI model construction or AI agent exploration, with multi-agent frameworks largely relying on LLMs. However, supervised models often outperform LLMs on specific tasks. It is imprudent to disregard these models in the LLM era. Therefore, we propose that admin agents should not only select agents but also choose models for various checks or evaluations of the agents’ outputs from multiple perspectives. Additionally, agents can utilize feedback from these models to refine their outputs further. Figure 1.2 presents an example. Imagine that we are now company managers preparing a speech script for an earnings conference call. We can utilize an LLM to generate the script by providing a set of references. Additionally, we can rely on AI agents to discuss various aspects, such as compliance and professionalism. However, it is also important to consider the market reaction to the speech script. Several models are specifically designed to predict market reactions based on earnings calls [ 14, 17, 18]. Given that these models perform better than agents, it is necessary to integrate them into the multi-agent framework. This is why we propose the concept of multi- scale model synergy. Furthermore, managers may need to anticipate the types of questions that professional analysts might ask [ 13], and models fine-tuned for this purpose should be adopted rather than relying solely on prompting LLMs. More discussions will be provided in Chap. 5. These are the reasons we believe that agent AI is a promising direction. In this book, we will share findings based on previous explorations and further discuss how to advance the concept of multi-scale model synergy.

📄 Page 18

References 7 1.4 Overview of the Book Chapter 2 connects our previous work with this book by providing a clearer under- standing of financial argument mining. This chapter also presents some of our explo- rations on the topic, and the results reveal the importance of incorporating argument- mining concepts into financial NLP. In Chap. 3, we begin the discussion on single- agent design and offer a survey of recent studies, covering topics such as multimodal agents, feature engineering and promotion, model editing, human annotation and feedback, and retrieval-augmented generation (RAG). Chapter 4 outlines the current state of multi-agent interaction development, sharing insights and findings related to behavior simulation, trading decision-making, and multimodal multi-agent inter- action. In Chap. 5, we discuss the proposed multi-scale model synergy and identify several open research questions. Chapter 6 explores additional use cases of argu- ment mining in finance and agent AI designed for finance. Finally, we highlight future research directions for agent AI in finance and conclude the book. References 1. Axtell, R. L., and Farmer, J. D. Agent-based modeling in economics and finance: Past, present, and future. Journal of Economic Literature (2022), 1–101. 2. Bollen, J., Mao, H., and Zeng, X. Twitter mood predicts the stock market. Journal of Computational Science 2, 1 (2011), 1–8. 3. Chen, C.-C., Huang, H.-H., and Chen, H.-H. Crowd View: Converting investors’ opinions into indicators. In IJCAI (2019), pp. 6500–6502. 4. Chen, C.-C., Huang, H.-H., and Chen, H.-H. Numeral attachment with auxiliary tasks. In Proceedings of the Forty-Second International ACM SIGIR Conference on Research and Development in Information Retrieval (2019), pp. 1161–1164. 5. Chen, C.-C., Huang, H.-H., and Chen, H.-H. From opinion mining to financial argument mining. Springer Nature, 2021. 6. Chen, C.-C., Huang, H.-H., Shiue, Y.-T., and Chen, H.-H. Numeral understanding in financial tweets for fine-grained crowd-based forecasting. In 2018 IEEE/WIC/ACM Interna- tional Conference on Web Intelligence (WI) (2018), IEEE, pp. 136–143. 7. Chen, C.-C., Huang, H.-H., Tsai, C.-W., and Chen, H.-H. CrowdPT: Summarizing crowd opinions as professional analyst. In The World Wide Web Conference (2019), pp. 3498– 3502. 8. Clement, M. B., and Tse, S. Y. Financial analyst characteristics and herding behavior in forecasting. The Journal of finance 60, 1 (2005), 307–341. 9. De Bondt, W. F., and Thaler, R. Does the stock market overreact? The Journal of finance 40, 3 (1985), 793–805. 10. Durante, Z., Huang, Q., Wake, N., Gong, R., Park, J. S., Sarkar, B., Taori, R., Noda, Y., Terzopoulos, D., Choi, Y., et al. Agent AI: Surveying the horizons of multimodal interaction. arXiv preprint arXiv:2401.03568 (2024). 11. Guo, T., Chen, X., Wang, Y., Chang, R., Pei, S., Chawla, N., Wiest, O., and Zhang, X. Large language model based multi-agents: A survey of progress and challenges. In 33rd International Joint Conference on Artificial Intelligence (IJCAI 2024) (2024), IJCAI. 12. Hong, S., Zhuge, M., Chen, J., Zheng, X., Cheng, Y., Wang, J., Zhang, C., Wang, Z., Yau, S. K. S., Lin, Z., et al. Metagpt: Meta programming for a multi-agent collabo- rative framework. In The Twelfth International Conference on Learning Representations.

📄 Page 19

8 1 Introduction 13. Juan, Y., Chen, C.-C., Huang, H.-H., and Chen, H.-H. Generating multiple questions from presentation transcripts: A pilot study on earnings conference calls. In Proceedings of the 16th International Natural Language Generation Conference (Prague, Czechia, Sept. 2023), C. M. Keet, H.-Y. Lee, and S. Zarrieß, Eds., Association for Computational Linguistics, pp. 449– 454. 14. Koval, R., Andrews, N., and Yan, X. Forecasting earnings surprises from confer- ence call transcripts. In Findings of the Association for Computational Linguistics: ACL 2023 (Toronto, Canada, July 2023), A. Rogers, J. Boyd-Graber, and N. Okazaki, Eds., Association for Computational Linguistics, pp. 8197–8209. 15. Li, J., Durmus, E., and Cardie, C. Exploring the role of argument structure in online debate persuasion. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP) (Online, Nov. 2020), Association for Computational Linguis- tics, pp. 8905–8912. 16. Park, J. S., O’Brien, J., Cai, C. J., Morris, M. R., Liang, P., and Bernstein, M. S. Generative agents: Interactive simulacra of human behavior. In Proceedings of the 36th Annual ACM Symposium on User Interface Software and Technology (New York, NY, USA, 2023), UIST ’23, Association for Computing Machinery. 17. Qin, Y., and Yang, Y. What you say and how you say it matters: Predicting stock volatility using verbal and vocal cues. In Proceedings of the Fifty-Seventh Annual Meeting of the Associa- tion for Computational Linguistics (Florence, Italy, July 2019), Association for Computational Linguistics, pp. 390–401. 18. Sang, Y., and Bao, Y. DialogueGAT: A graph attention network for financial risk prediction by modeling the dialogues in earnings conference calls. In Findings of the Association for Computational Linguistics: EMNLP 2022 (Abu Dhabi, United Arab Emirates, Dec. 2022), Y. Goldberg, Z. Kozareva, and Y. Zhang, Eds., Association for Computational Linguistics, pp. 1623–1633. 19. Skala, D. Overconfidence in psychology and finance-an interdisciplinary literature review. Bank I kredyt, 4 (2008), 33–50. 20. Toulmin, S. E. The Uses of Argument. Cambridge University Press, 2003. 21. Wachsmuth, H., Al Khatib, K., and Stein, B. Using argument mining to assess the argumentation quality of essays. In Proceedings of COLING 2016, the Twenty-Sixth Interna- tional Conference on Computational Linguistics: Technical Papers (2016), pp. 1680–1691. Open Access This chapter is licensed under the terms of the Creative Commons Attribution 4.0 International License (http://creativecommons.org/licenses/by/4.0/), which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license and indicate if changes were made. The images or other third party material in this chapter are included in the chapter’s Creative Commons license, unless indicated otherwise in a credit line to the material. If material is not included in the chapter’s Creative Commons license and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder.

📄 Page 20

Chapter 2 Financial Argument Mining This chapter focuses on financial argument mining. In Sect. 2.1, we first recap the concept of the argument structure of financial opinions, which we discussed in our previous work [ 6]. We also discuss the statistics of argument structures of company managers and professional analysts based on the annotated data. In Sect. 2.2, we introduce an extended concept, forward-looking argument mining, by presenting the ideas of “scenario” and “impact duration.” The demonstration of how to apply the concept of forward-looking argument mining for downstream tasks is provided in Sect. 2.3. Finally, we conclude this chapter with a summary in Sect. 2.4. 2.1 Argument Structure In Sects. 2.2 and 2.3 of our previous book [ 6], we introduced the concept of convert- ing raw opinions from investors into an argument structure. This encompasses both the internal structure of a single opinion and the relationships among multiple opin- ions. Rather than simply presenting the raw analysis from a professional analyst, we propose breaking it down into distinct argument units and further understanding it in a structured form. These argument units can be categorized into claims and premises, and the final recommendation (e.g., Overweight) is referred to as the main claim. Additionally, the connections between argument units vary in weight and quality. Some reasons may strongly support the claim, while others may be weaker due to the lack of a causal relationship. The overall quality of a claim is also influenced by the premises and inferences supporting it. These concepts have been previously explored and discussed. Recently, we have constructed datasets based on these concepts, including stock analysis reports [ 17] and earnings conference calls [ 3]. Although we identified four primary sources for financial NLP [ 6]—managers, professionals, social media users, and journals— extracting premises and claims from the latter two is challenging. Social media users, © The Author(s) 2025 C.-C. Chen and H. Takamura, Agent AI for Finance, SpringerBriefs in Intelligent Systems, https://doi.org/10.1007/978-3-031-94687-5_2 9