請用此 Handle URI 來引用此文件：
http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/102663完整後設資料紀錄
| DC 欄位 | 值 | 語言 |
|---|---|---|
| dc.contributor.advisor | 陳尚澤 | zh_TW |
| dc.contributor.advisor | Shang-Tse Chen | en |
| dc.contributor.author | 林俊霆 | zh_TW |
| dc.contributor.author | Jun-Ting Lin | en |
| dc.date.accessioned | 2026-07-15T16:14:19Z | - |
| dc.date.available | 2026-07-16 | - |
| dc.date.copyright | 2026-07-15 | - |
| dc.date.issued | 2026 | - |
| dc.date.submitted | 2026-07-07 | - |
| dc.identifier.citation | [1] Harry Markowitz. Portfolio selection. *The Journal of Finance*, 7(1):77–91, 1952.
[2] Fischer Black and Robert Litterman. Global portfolio optimization. *Financial Analysts Journal*, 48(5):28–43, 1992. [3] Shijie Wu, Ozan Irsoy, Steven Lu, Vadim Dabravolski, Mark Dredze, Sebastian Gehrmann, Prabhanjan Kambadur, David Rosenberg, and Gideon Mann. BloombergGPT: A large language model for finance, 2023. [4] Yunchuan Sun, Lu Liu, Ying Xu, Xiaoping Zeng, Yufeng Shi, Haifeng Hu, Jie Jiang, and Ajith Abraham. Alternative data in finance and business: emerging applications and theory analysis (review). *Financial Innovation*, 10(1):127, 2024. [5] Paul C. Tetlock. Giving content to investor sentiment: The role of media in the stock market. *The Journal of Finance*, 62(3):1139–1168, 2007. [6] Allen H. Huang, Hui Wang, and Yi Yang. FinBERT: A large language model for extracting information from financial text. *Contemporary Accounting Research*, 40(2):806–841, 2023. [7] Pinyu Chen, Zois Boukouvalas, and Roberto Corizzo. A deep fusion model for stock market prediction with news headlines and time series data. *Neural Computing and Applications*, 36(34):21229–21271, 2024. [8] Kemal Kirtac and Guido Germano. Sentiment trading with large language models. *Finance Research Letters*, 62:105227, 2024. [9] Wei Ding, Zhennan Chen, Hanpeng Jiang, Yuanguo Lin, and Fan Lin. Trend-heuristic reinforcement learning framework for news-oriented stock portfolio management. In *Proceedings of the IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)*, pages 5120–5124, 2024. [10] Chun-Yu Marvin Chuang, Jun-Ting Lin, Yung-Yaw Chen, Yi-Ren Yeh, Chun-Yuan Huang, and Jyh-Shing Roger Jang. VisionTrader: Portfolio optimization with deep reinforcement learning and Vision Transformer. Manuscript submitted for publication. Department of Computer Science and Information Engineering, National Taiwan University, 2026. [11] John Arevalo, Thamar Solorio, Manuel Montes-y-Gómez, and Fabio A. González. Gated multimodal units for information fusion, 2017. [12] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Łukasz Kaiser, and Illia Polosukhin. Attention is all you need. In *Advances in Neural Information Processing Systems*, volume 30, pages 5998–6008, 2017. [13] Sébastien Maillard, Thierry Roncalli, and Jérôme Teïletche. The properties of equally weighted risk contribution portfolios. *The Journal of Portfolio Management*, 36(4):60–70, 2010. [14] Richard S. Sutton and Andrew G. Barto. *Reinforcement Learning: An Introduction*. MIT Press, 2nd edition, 2018. [15] Ronald J. Williams. Simple statistical gradient-following algorithms for connectionist reinforcement learning. *Machine Learning*, 8(3–4):229–256, 1992. [16] Zhengyao Jiang, Dixing Xu, and Jinjun Liang. A deep reinforcement learning framework for the financial portfolio management problem, 2017. [17] Zhicheng Wang, Biwei Huang, Shikui Tu, Kun Zhang, and Lei Xu. DeepTrader: A deep reinforcement learning approach for risk-return balanced portfolio management with market conditions embedding. In *Proceedings of the AAAI Conference on Artificial Intelligence*, volume 35, pages 643–650, 2021. [18] Bo Pang, Lillian Lee, and Shivakumar Vaithyanathan. Thumbs up? Sentiment classification using machine learning techniques. In *Proceedings of the 2002 Conference on Empirical Methods in Natural Language Processing (EMNLP)*, pages 79–86, 2002. [19] Tim Loughran and Bill McDonald. When is a liability not a liability? Textual analysis, dictionaries, and 10-Ks. *The Journal of Finance*, 66(1):35–65, 2011. [20] 張津挺. 中文財務情緒字典建構與其在財務新聞分析之應用. Master's thesis, 臺北市立大學, 2015. [21] Hongyang Yang, Xiao-Yang Liu, and Christina Dan Wang. FinGPT: Open-source financial large language models, 2023. [22] 洪嘉濃. 整合大型語言模型情感分析與機器學習之股票漲跌預測方法. Master's thesis, 國立臺灣大學, 2025. [23] Tadas Baltrušaitis, Chaitanya Ahuja, and Louis-Philippe Morency. Multimodal machine learning: A survey and taxonomy. *IEEE Transactions on Pattern Analysis and Machine Intelligence*, 41(2):423–443, 2019. | - |
| dc.identifier.uri | http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/102663 | - |
| dc.description.abstract | 隨著大型語言模型的崛起，新聞等非結構化資訊成為量化投資中可提取alpha的訊號來源。然而，將新聞情緒整合至多股票深度強化學習（Deep ReinforcementLearning，DRL）投資組合最佳化的研究仍屬稀缺。
本研究以VisionTrader 為基礎架構，於台股元大台灣50 成分股上探索新聞情緒融入多股票DRL 投資組合框架的可行性，圍繞三個遞進目的展開：（一）檢視MSU（Market Score Unit）的貢獻以確立基準架構；（二）比較不同情緒整合配置在不同覆蓋率股票池的效能，並比較四種融合機制以找出最佳整合方式；（三）跨八個時間區段驗證穩健性。 研究結果表明，MSU 在月頻設定下貢獻有限，可安全移除。融合效益高度依賴股票池新聞覆蓋率，高覆蓋率子集可獲顯著提升，混合覆蓋率池則反而衰退；四種融合機制中Gate 表現最佳。跨期分析顯示，純價量模型在八個區段中有六個優於0050；新聞融合相對於純價量的優勢具市場環境依賴性——多頭市場帶來顯著增益，但平盤高雜訊市場此優勢消失。 本研究揭示新聞覆蓋率為融合效益的關鍵調節因素，填補新聞情緒與多股票DRL 投資組合整合的文獻缺口，並為從業者提供實務指引。 | zh_TW |
| dc.description.abstract | With the rise of large language models, unstructured information such as financial news has emerged as a valuable source of alpha signals in quantitative investing. However, research integrating news sentiment into multi-stock deep reinforcement learning (DRL) portfolio optimization remains scarce.
This study takes VisionTrader as the backbone architecture and explores the feasibility of incorporating news sentiment into a multi-stock DRL portfolio framework on Taiwan's Yuanta Taiwan 50 ETF (0050) constituents, organized around three progressive objectives: (1) examining the contribution of the Market Score Unit (MSU) to establish a baseline architecture; (2) comparing different sentiment integration configurations across stock pools with varying news coverage rates, and comparing four fusion mechanisms to identify the optimal integration method; (3) verifying cross-period robustness across eight rolling time windows. Results show that MSU contributes marginally under monthly settings and can be safely removed. Fusion effectiveness is highly dependent on news coverage rate: high-coverage subsets achieve significant improvements, while mixed-coverage pools suffer performance degradation; among the four fusion mechanisms, Gate performs best. Cross-period analysis shows that the pure price-volume model outperforms 0050 in six out of eight segments; the advantage of news fusion over the pure price-volume model is market-environment-dependent—yielding significant gains in bull markets but diminishing in sideways, high-noise markets. This study reveals news coverage rate as a critical moderating factor for fusion effectiveness, fills the literature gap at the intersection of news sentiment and multi-stock DRL portfolio optimization, and provides practical guidelines for practitioners. | en |
| dc.description.provenance | Submitted by admin ntu (admin@lib.ntu.edu.tw) on 2026-07-15T16:14:19Z No. of bitstreams: 0 | en |
| dc.description.provenance | Made available in DSpace on 2026-07-15T16:14:19Z (GMT). No. of bitstreams: 0 | en |
| dc.description.tableofcontents | Verification Letter from the Oral Examination Committee i
Acknowledgements iii 摘要 v Abstract vii Contents ix List of Figures xv List of Tables xvii Chapter 1 Introduction 1 1.1 Research Background 1 1.2 Research Motivation 3 1.3 Research Objectives 4 1.4 Thesis Structure 5 Chapter 2 Literature Review 7 2.1 Portfolio Optimization Theory 7 2.1.1 Markowitz Mean-Variance Model 7 2.1.2 Further Developments and Limitations of Traditional Methods 8 2.2 Reinforcement Learning for Portfolio Management 9 2.2.1 Fundamentals of Reinforcement Learning 9 2.2.2 Policy Gradient Methods 10 2.2.3 Representative Research on Deep Reinforcement Learning for Portfolio Management 11 2.3 DeepTrader, VisionTrader, and TrendTrader 11 2.3.1 DeepTrader: Dual-Unit Architecture for RL Portfolio Management 12 2.3.2 VisionTrader: Enhanced Spatiotemporal Modeling with ViT 13 2.3.3 TrendTrader: A Reinforcement Learning Portfolio Framework with News Sentiment 13 2.4 News Sentiment Analysis: Empirical Foundations and Methodological Evolution 15 2.4.1 Empirical Research on the Impact of News Sentiment on Stock Markets 15 2.4.2 Evolution of Sentiment Indicator Construction Methods 15 2.5 Financial Text Sentiment Analysis: Discriminative and Generative Models 17 2.5.1 BERT Series: Discriminative Models for the Financial Domain 17 2.5.2 Applications of Generative LLMs in Financial Text 18 2.6 Multimodal Fusion: Integrating Price-Volume Data and News Sentiment 19 2.6.1 Related Research on Integrating Structured and Unstructured Information 19 2.6.2 Fusion Timing and Feature-Level Methods in Multimodal Fusion 20 Chapter 3 Methodology 23 3.1 System Architecture 23 3.1.1 ASU Architecture Overview 25 3.1.2 Output and Stock Selection 25 3.2 News Sentiment Feature Extraction 26 3.2.1 News Preprocessing Pipeline 26 3.2.2 FinBERT News Sentiment Score (Path A) 27 3.2.3 FinBERT News Sentiment Embedding (Path B) 28 3.2.4 Temporal Aggregation 29 3.3 Feature Fusion Strategy 30 3.3.1 Fusion Architecture Overview 30 3.3.2 Concatenation 31 3.3.3 Addition 31 3.3.4 Gating 32 3.3.5 Cross-Attention 33 3.3.6 Comparison of Fusion Methods 34 3.4 Reinforcement Learning Training Setup 34 3.4.1 Problem Formalization (MDP) 34 3.4.2 Stock Selection Weight Calculation and Capital Allocation 36 3.4.3 Trade Execution Flow 37 3.4.4 Policy Optimization Objective 38 Chapter 4 Experimental Design 39 4.1 Dataset 39 4.1.1 Stock Data 39 4.1.2 News Data 40 4.1.3 News Preprocessing 43 4.2 Experimental Settings 44 4.2.1 Data Splitting 44 4.2.2 Baseline Models 45 4.2.3 Evaluation Metrics 46 4.2.4 Implementation Details 47 Chapter 5 Experimental Results and Discussion 51 5.1 Baseline Performance 51 5.1.1 Experimental Configuration 51 5.1.2 Performance Comparison 52 5.1.3 Cumulative Wealth Curves 53 5.1.4 Summary 54 5.2 MSU Removal Experiment 54 5.2.1 Experimental Configuration 54 5.2.2 Analysis of MSU Output ρ Characteristics 55 5.2.3 Mixed Pool (47 Stocks) Performance Comparison 56 5.2.4 High-Coverage Pool (22 Stocks) Performance Comparison 57 5.2.5 Training Cost Comparison 58 5.2.6 Summary 59 5.3 ASU News Fusion Experiments 59 5.3.1 Experimental Configuration 59 5.3.2 Mixed Pool (47 Stocks) 60 5.3.3 High-Coverage Pool (22 Stocks) 62 5.3.4 Summary 63 5.4 Comparison of Different Fusion Methods 64 5.4.1 Experimental Configuration 64 5.4.2 Performance Comparison 65 5.4.3 Summary 67 5.5 Robustness Analysis 68 5.5.1 Recent Rolling Window Validation 68 5.5.2 Early Rolling Window Validation 71 5.5.3 Validation Period Sensitivity Analysis 73 5.5.4 Summary 75 Chapter 6 Conclusions and Future Research Directions 77 6.1 Research Conclusions 77 6.2 Research Contributions 79 6.3 Research Limitations 82 6.4 Future Research Directions 84 References 87 Appendix A — News Coverage Details for 47 Constituent Stocks 91 | - |
| dc.language.iso | en | - |
| dc.subject | 深度強化學習 | - |
| dc.subject | 投資組合最佳化 | - |
| dc.subject | 新聞情緒分析 | - |
| dc.subject | 多模態融合 | - |
| dc.subject | FinBERT | - |
| dc.subject | 台灣股市 | - |
| dc.subject | Deep Reinforcement Learning | - |
| dc.subject | Portfolio Optimization | - |
| dc.subject | News Sentiment Analysis | - |
| dc.subject | Multimodal Fusion | - |
| dc.subject | FinBERT | - |
| dc.subject | Taiwan Stock Market | - |
| dc.title | 新聞特徵融合於多模態深度強化學習之投資組合最佳化：以台灣股市為例 | zh_TW |
| dc.title | News Feature Fusion in Multimodal Deep Reinforcement Learning for Portfolio Optimization: Evidence from the Taiwan Stock Market | en |
| dc.type | Thesis | - |
| dc.date.schoolyear | 114-2 | - |
| dc.description.degree | 碩士 | - |
| dc.contributor.coadvisor | 張智星 | zh_TW |
| dc.contributor.coadvisor | Jyh-Shing Roger Jang | en |
| dc.contributor.oralexamcommittee | 陳永耀;葉倚任 | zh_TW |
| dc.contributor.oralexamcommittee | Yung-Yaw Chen;Yi-Ren Yeh | en |
| dc.subject.keyword | 深度強化學習; 投資組合最佳化; 新聞情緒分析; 多模態融合; FinBERT; 台灣股市 | zh_TW |
| dc.subject.keyword | Deep Reinforcement Learning; Portfolio Optimization; News Sentiment Analysis; Multimodal Fusion; FinBERT; Taiwan Stock Market | en |
| dc.relation.page | 92 | - |
| dc.identifier.doi | 10.6342/NTU202601633 | - |
| dc.rights.note | 同意授權(全球公開) | - |
| dc.date.accepted | 2026-07-08 | - |
| dc.contributor.author-college | 電機資訊學院 | - |
| dc.contributor.author-dept | 資訊網路與多媒體研究所 | - |
| dc.date.embargo-lift | 2026-07-16 | - |
| 顯示於系所單位： | 資訊網路與多媒體研究所 | |
文件中的檔案：
| 檔案 | 大小 | 格式 | |
|---|---|---|---|
| ntu-114-2.pdf | 5.11 MB | Adobe PDF | 檢視/開啟 |
系統中的文件，除了特別指名其著作權條款之外，均受到著作權保護，並且保留所有的權利。