融合深度神經網路與深層模糊孿生支持向量機於股價預測

A Stock Closing Price Prediction Model based on Deep Neural Networks and Deep Fuzzy Twin Support Vector Machine

龔千芬(Chien-Feng Kung) ； 郝沛毅(Pei-Yi Hao)

《資訊管理學報》 29卷4期 (2022/10) Pp. 303-333

摘要

股價預測是橫跨金融與計算機科學領域的經典預測問題，由於成功預測股價的潛在好處，它吸引一代又一代的學者與投資者從不同的角度、無數的學理、眾多的投資策略和不同的實踐經驗來開發各種預測方法。股價預測的困難癥結點在於影響股票漲跌的因素太多。股市波動通常是由熱門新聞推動的，社群媒體上的推文則是反映了新聞事件的熱度，以及投資者對於該事件的態度。因此，分析社群推文的文字資訊與股價技術指數的數值資料，為能夠幫助於我們預測未來的股價變化。由於股價受到眾多因素影響，很難通過簡單的模型進行預測。深度學習擁有優異的特徵學習能力，支持向量機則是擁有優異的推理能力，本研究結合兩者的優點。本研究提出一個混和深度模型來自動學習重要的特徵，該混和深度模型是由卷積神經網絡（CNN）、雙向長短期記憶（BiLSTM）與注意力機制（AM）組成。CNN用於擷取輸入數據的位置不變特徵，BiLSTM則是提取長時間依賴性的特徵，AM用於捕捉過去不同時間特徵狀態對股票收盤價的影響，以提高預測的正確率。接著，本研究將擷取得到的特徵餵給模糊孿生支持向量機來建立最佳的股價預測模型，並且透過轉移學習理論建立嶄新的深層模糊孿生支持向量機。本研究在台積電股票的預測正確率最高為76.9667％，友達股票的預測正確率最高為87.0856％，與經典的股價預測模型相比，本研究所提出的方法的預測正確率明顯優於最先進的股價預測模型。

關鍵字

股價預測 ； 模糊支持向量機 ； 卷積神經網路 ； 雙向長短期記憶體 ； 注意力機制

並列摘要

Stock closing price prediction is an important problem in the intersection of computer science and finance. Due to the potential advantages of stock closing price prediction, it attracts generation after generation of investors as well as scholars to continuously develop various prediction methods from different perspectives, a multitude of investment strategies, different practical experiences, and a myriad of theories. The stock closing price is often affected by hot news, and the tweets related to news reflect the heat of the breaking news, as well as the sentiment of investors towards the breaking news. Consequently, analyzing tweets and historical stock market indices may help us to predict future price changes. Since the stock price is affected by many factors, it is difficult to predict through a simple model. Deep learning methods have the advantage of learning features. Support vector machines have the advantage of generalizing very well. This paper combines the advantages of both models. Therefore, this paper develops a hybrid deep model to automatically learn important features. This deep model is composed of convolutional neural networks (CNN), bi-directional long short-term memory (BiLSTM), and attention mechanism (AM). CNN is utilized to discover the time-invariant features of the input data. BiLSTM is used to extract time dependency features. AM is used to extract the influence of feature states on the stock closing price at different times in the past to improve the classification performance. The obtained feature is fed to a deep fuzzy support vector machine to build an optimal stock prediction model. As for this study, the highest forecast accuracy is 76.9667% and 87.0856% for Taiwan Semiconductor Manufacturing Corporation and AU Optronics Corporation, respectively. When compared with previous prediction models, the method proposed in this study is significantly better than the state-of-the-art support vector machine and deep learning models.

並列關鍵字

Stock prediction ； Fuzzy support vector machine ； Convolutional neural networks ； Bi-directional long short-term memory ； Attention mechanism

參考文獻

Xu, Y. & Cohen, S. B. (2018). Stock movement prediction from tweets and historical prices, Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics, 1970-1979.

Yan, Y. & Yang, D. (2021a). A stock trend forecast algorithm based on deep neural networks, Scientific Programming, 2021(7510641), 1-7.

Yan, X., Weihan, W., & Chang, M. (2021b). Research on financial assets transaction prediction model based on LSTM neural network, Neural Computing and Applications, 33, 257-270.

Yun, H., Sim, G., & Seok, J. (2019). Stock prices prediction using the title of newspaper articles with Korean natural language processing, International Conference on Artificial Intelligence in Information and Communication (ICAIIC), 019-021.

Yun, K. K., Yoon, S. W., & Won, D. (2021). Prediction of stock price direction using a hybrid GA-XGBoost algorithm with a three-stage feature engineering process, Expert Systems with Applications, 186, 115716.