| 摘要: | 目前股價預測研究多是基於技術面的數據資料，然而隨著網路社群的發展，消息面的發酵會影響投資人的決策和短期股價走向。因此，如何萃取消息面的文本資訊，進而能同時考量技術面和消息面的影響，將是未來提升股價預測準確性的重要研究方向。
本研究主要是利用近年來改良的動態與靜態詞嵌入方法，預測社群媒體文本的情感分數後，結合技術面資料後，透過深度學習分類器進行股價預測。而在深度學習分類器中，在情感分析方法的部分是根據近年來在文本序列上十分熱門的Transformer架構進行不同組態的改良搭配，以找出最適用於情感分析的Transformer組態。而在股價預測方面，則是提出一個結合混合式模型中的CNN，與Transformer架構中的Multi-head Self-Attention機制，搭配LSTM進行改良的分類器(Multi-head Self-Attention with Convolutional neural network LSTM, MAC-LSTM)，加強重要訊息的關注，以更準確掌握股票漲跌的趨勢。
本研究蒐集兩個社群媒體文本(Twitter、Stocktwit)的資料並以亞馬遜公司(AMZN)與網飛公司(NFLX)為預測標的，搭配不同的情感分析方法(辭典法、API、動靜態詞嵌入情感分析方法)，並以三種深度分類器，ATLSTM、CNN-LSTM和MAC-LSTM進行股價預測，探討其何種組合最適用於股價預測。最終，本實驗發現以內容面改良的動態詞嵌入情感分析方法finBERT，搭配深度學習分類器MAC-LSTM能最精準的預測股票的未來漲跌趨勢，其準確率最高達71.70%，為最佳組合。
;At present, most research on stock price prediction is based on technical data. However, With the development of social media, investors can easily receive numerous information so that they will make different decisions, and thus cause the short-term inflation of the stock price. Therefore, considering both text and technical data will be an important direction for improving the accuracy of stock price prediction.
This research is mainly discussed the efficiency about using several improved dynamic and static word embedding methods to conduct the sentiment scores of social media texts, and then combine them with structured data to predict stock trend by a deep learning classifier. In the deep learning classifier, we figure out the best Transformer configuration which called Enhanced Transformer for sentiment analysis. In the area of stock price prediction, we propose an improved classifier (Multi-head Self-Attention with Convolutional Neural Network LSTM, MAC) by combining CNN in hybrid model and Multi-head Self-Attention mechanism in Transformer architecture. MAC-LSTM) to enhance the attention of important information to grasp the trend of stock rise and fall more accurately.
This research collects data from two kinds of social media text data (Twitter and Stocktwit) and uses Amazon (AMZN) and Netflix (NFLX) as the predict targets, with different sentiment analysis methods (dictionary method, API, and dynamic word embedding sentiment analysis method), and three deep classifiers, ATLSTM, CNN-LSTM, and MAC-LSTM, for investigating which combination is most suitable for stock prediction. At the end of the research, we found that the improved dynamic word embedding sentiment analysis method finBERT with the stock prediction classifier MAC-LSTM can predict the future stock trend with the best combination of 71.70% accuracy. |
