本國股市波動易受投資者之情緒變化，而投資者則會因為新聞而導致投資決策的改變。本文透過爬蟲技術取得近十年華爾街日報科技版的每日新聞資料，結合深度學習模型進行情感分析，期望能夠幫助投資者更好地了解股票市場的變化。現有的情感分析大多缺少金融詞彙的訓練，故本文利用 FinBERT 預訓練模型去學習金融新聞語義的特徵，解決上述問題並針對臺灣及美國的科技類股指數的市場報酬率進行分析。研究發現，在科技領域的股市中，華爾街日報只會影響臺灣股價，並不會影響美國股價，故華爾街日報在臺灣股市之中可視為領先指標，反之，在美國股市中為落後指標，另外，在臺灣科技股股市中， COVID19 事件發生後會增加股市的波動性，對市場的可解釋力也上升。除此之外，由華爾街日報得出的情緒分數變動量和台灣及美國科技股市中之報酬率存在邊際效果遞減，故本研究認為情緒因子可以加入其他迴歸模型以得到更好的預測能力。

The domestic stock market is highly sensitive to changes in investor sentiment, and investors often make investment decisions based on news. This paper uses web crawling technology to obtain daily news data from the technology section of The Wall Street Journal over the past decade. The author then use a deep learning model for sentiment analysis to better understand changes in the stock market. Existing sentiment analysis methods mostly lack training on financial vocabulary, so this study uses the FinBERT pre-training model to learn the semantic features of financial news, addressing this issue and analyzing the market returns of Taiwan and US technology sector indices. The study found that The Wall Street Journal only affects Taiwan's stock prices in the technology sector, and doesn't affect US stock prices. Therefore, The Wall Street Journal can be considered a leading indicator in the Taiwan stock market, whereas it is a lagging indicator in the US stock market. After the COVID-19 event, not only did it increase the volatility of the stock market, but also the market's interpretability. In addition, the marginal effect of the sentiment score volatility derived from The Wall Street Journal on the returns of both Taiwan and US markets exhibits diminishing marginal returns, so this study believes that sentiment factors can be added to other regression models to achieve better predictive ability.

目錄
1 緒論 1
1.1 研究背景 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 研究動機 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.3 研究目的 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.4 研究流程 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
2 文獻探討 6
2.1 基於深度學習模型進行文本分類 . . . . . . . . . . . . . . . . . . . . . 6
2.2 基於深度學習模型進行情緒分析 . . . . . . . . . . . . . . . . . . . . . 8
2.3 情緒分析與市場的可預測性 . . . . . . . . . . . . . . . . . . . . . . . 8
3 研究方法 11
3.1 資料蒐集 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
3.1.1 文本數據 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
3.1.2 指數數據 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
3.2 文本處理 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
3.2.1 篩選資料 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
3.2.2 整理資料 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
3.3 模型運用 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
3.3.1 FinBERT 模型 . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
3.3.2 Transformer 模型 . . . . . . . . . . . . . . . . . . . . . . . . . . 17
3.4 情緒分數計算 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
3.4.1 方法一：設定三種情緒標籤分數之參數 . . . . . . . . . . . . 18
3.4.2 方法二：選定單一標籤情緒分數 . . . . . . . . . . . . . . . . 19
3.5 情緒模型建立 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
3.5.1 模型一：基本迴歸模型 . . . . . . . . . . . . . . . . . . . . . . 20
3.5.2 模型二：情緒分數變動量迴歸模型 . . . . . . . . . . . . . . . 21
3.5.3 模型三：在模型一中加入大盤控制因素 . . . . . . . . . . . . 21
3.5.4 模型四：在模型二中加入大盤控制因素 . . . . . . . . . . . . 22
3.6 動態分析 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
4 實證結果與分析 24
4.1 迴歸分析 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
4.1.1 方法一：設定三種情緒標籤分數之參數 . . . . . . . . . . . . 24
4.1.2 方法二：選定單一標籤情緒 . . . . . . . . . . . . . . . . . . . 31
4.2 動態分析之圖形呈現 . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
4.2.1 比較每日情緒分數跟半導體類指數的市場報酬率 . . . . . . . 36
4.2.2 比較每日情緒分數變動量跟半導體類指數的市場報酬率 . . . 38
4.2.3 比較每日情緒分數跟電子類指數的市場報酬率 . . . . . . . . 39
4.2.4 比較每日情緒分數變動量跟電子類指數的市場報酬率 . . . . 40
4.2.5 比較 COVID19 事件發生之影響 . . . . . . . . . . . . . . . . . 41
5 結論與建議 46
5.1 結論 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
5.2 建議 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
參考文獻 48

參考文獻
[1] Dogu Araci. Finbert: Financial sentiment analysis with bert. https://github.com/ProsusAI/finBERT, 2019.
[2] Dogu Araci. Finbert: Financial sentiment analysis with pre-trained language models. arXiv preprint arXiv:1908.10063, 2019.
[3] Malcolm Baker and Jeffrey Wurgler. Investor sentiment in the stock market. Journal of economic perspectives, 21(2):129–152, 2007.
[4] Luc Bellintani and Junior Sophistor. The effect of news on intra-day stock prices & their volatility. Student Economic Review, 33:111–121, 2019.
[5] Charles W Calomiris and Harry Mamaysky. How news and its context drive risk and returns around the world. Journal of Financial Economics, 133(2):299–336, 2019.
[6] Nicola Capuano, Luca Greco, Pierluigi Ritrovato, and Mario Vento. Sentiment analysis for customer relationship management: An incremental learning approach. Applied Intelligence, 51(6):3339–3352, 2021.
[7] Sanjiv R Das and Mike Y Chen. Yahoo! for amazon: Sentiment extraction from small talk on the web. Management science, 53(9):1375–1388, 2007.
[8] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pretraining of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018.
[9] Xing Fang and Justin Zhan. Sentiment analysis using product review data. Journal of Big Data, 2(1):1–14, 2015.
[10] Paul Glasserman and Harry Mamaysky. Does unusual news forecast market stress? Journal of Financial and Quantitative Analysis, 54(5):1937–1974, 2019.
[11] James B Heaton, Nick G Polson, and Jan Hendrik Witte. Deep learning for finance: deep portfolios. Applied Stochastic Models in Business and Industry, 33(1):3–12, 2017.
[12] Steven L Heston and Nitish Ranjan Sinha. News vs. sentiment: Predicting stock returns from news stories. Financial Analysts Journal, 73(3):67–83, 2017.
[13] Jeremy Howard and Sebastian Ruder. Fine-tuned language models for text classification. CoRR, abs/1801.06146, 2018.
[14] Pekka Malo, Ankur Sinha, Pyry Takala, Pekka J. Korhonen, and Jyrki Wallenius. Good debt or bad debt: Detecting semantic orientations in economic texts. CoRR, abs/1307.5336, 2013.
[15] Federico Neri, Carlo Aliprandi, Federico Capeci, Montserrat Cuadros, and Tomas By. Sentiment analysis on social media. In 2012 IEEE/ACM international conference
on advances in social networks analysis and mining, pages 919–926. IEEE, 2012.
[16] Matthew E. Peters, Mark Neumann, Mohit Iyyer, Matt Gardner, Christopher Clark, Kenton Lee, and Luke Zettlemoyer. Deep contextualized word representations. CoRR, abs/1802.05365, 2018.
[17] Jie Ren, Hang Dong, Balaji Padmanabhan, and Jeffrey V Nickerson. How does social media sentiment impact mass media sentiment? a study of news in the financial markets. Journal of the Association for Information Science and Technology, 72(9):1183–1197, 2021.
[18] Robert P Schumaker, Yulei Zhang, Chun-Neng Huang, and Hsinchun Chen. Evaluating sentiment in financial news articles. Decision Support Systems, 53(3):458–464, 2012.
[19] Ankur Sinha, Satishwar Kedas, Rishu Kumar, and Pekka Malo. Sentfin 1.0: Entityaware sentiment analysis for financial news. Journal of the Association for Information Science and Technology, 2022.
[20] Wataru Souma, Irena Vodenska, and Hideaki Aoyama. Enhanced news sentiment analysis using deep learning methods. Journal of Computational Social Science, 2(1):33–46, 2019.
[21] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017.
[22] Wikipedia contributors. The wall street journal — Wikipedia, the free encyclopedia, 2023. [Online; accessed 2-May-2023].
[23] Chenrui Zhang. Research on the correlation between text emotion mining and stock market based on deep learning. arXiv preprint arXiv:2205.06675, 2022.