| 摘要: | 隨著科技的發展，投資者可以透過多種渠道，獲得金融商品的最新市場情報，增加整理分析資訊的時間成本，而難以透過自主分析準確判斷合適的投資時機，因此，如何掌握不同面向的股票買賣決策資訊，成為了獲利的重大關鍵。
本研究針對文本的內容主要提出利用靜態與動態詞嵌入，透過機器學習與深度學習分類器建立情感分析模型，預測社群媒體之文章內文的情感分數，再結合結構化資料，進行股票預測。過往的股票領域文本的情感分析研究中，雖然有許多研究針對股票相關文本使用詞嵌入方法建立情感分析模型，分辨其情感意涵，用以預測股票未來漲跌趨勢，但卻未有將此方法與其他情感方法加以比較，判斷不同情感分析方法在股票預測中的適用性，故本研究將所建構之詞嵌入情感分析模型，與傳統的辭典法和情感分析套件 Vader 進行比較，找尋有利於股票預測之情感分析方法。
本研究收集三個社群媒體文本數據(Twitter、Instagram、Facebook)，搭配不同的情感分析方法(辭典法、API、靜態詞嵌入情感分析模型)，並以四種不同的分類器，Random Forest、Na?ve Bayes、LSTM、PF-LSTM 進行分類，期望了解不同情感分析方法對於股票預測的作用。最終，本實驗發現動態詞嵌入 BERT 搭配 SO 的情感分析模型，可判別出社群媒體文本中所隱藏的情感意涵且對於股票漲跌預測結果具有正面的影響，準確率最高達 70.15%，為最佳情感分析模型。;With the development of technology, investors can obtain the latest market intelligence of financial products through multiple channels, and it also increase the time cost of analyzing information. It is difficult to accurately determine the appropriate investment timing through independent analysis. Therefore, how to master different aspects of stocks Information on buying and selling decisions has become a significant key to profit.
This research mainly proposes the use of static and dynamic word embedding for the content of the text, and the establishment of sentiment analysis models through machine learning and deep learning classifiers to predict the sentiment scores of articles in social media, and then combine structured data to make stock predictions. In the past research on sentiment analysis of texts, although many researches used word embedding methods to establish sentiment analysis models for stock-related texts to distinguish their sentimental meanings to predict the future rise and fall trends of stocks, but they didn’t use this method compare with other sentiment methods to judge the applicability of different sentiment analysis methods in stock forecasting. Therefore, this research will construct word representation sentiment analysis
model is compared with traditional dictionary methods and Vader API to find sentiment analysis methods that are conducive to stock forecasting.
This study collected three social media text data (Twitter, Instagram, Facebook), combined with different sentiment analysis methods (dictionary, API, static word embedding sentiment analysis model), and four classifiers, Random Forest , Na?ve Bayes, LSTM, PF-LSTM to classify, expect to understand the effect of different sentiment analysis methods on stock prediction. In the end, this experiment found Bidirectional Encoder Representations from Transformers and Sentiment Orientation emotional analysis model can identify the emotional meaning in the social media and has a positive impact on the stock price prediction. The accuracy rate is as high as 70.15%, which is the best sentiment analysis model. |
