無論在學術界或是業界,與股票市場趨勢預測相關的研究論文其數目是不勝枚舉,足以見得股市預測之重要性,本論文運用深層類神經學習網路的學習架構,並搭配著Long Short Term Memory、Embedding Layer之及Dropout等深度學習網路常見的網路結構建立出深度學習模型,並以台灣股市中544檔上市上櫃股票的最高點、最低點、開盤價、收盤價、成交量、融資餘額、融券餘額、三大法人(外資、投信、自營商)買賣超張數共十項特徵值,接著外加大盤最高點、最低點、開盤價、收盤價、成交量五項特徵值,合併共十五項特徵值向量欲預測一般市場買賣經驗定義獲利比例超過5%且可能損失風險小於3%之股票,以作為經濟發展與市場投資之參考。 除了以Long Short Term Memory網路結構建立模型外,本論文也嘗試以Convolutional Neural Network網路結構建立模型並比較其差異,最後利用視覺化分析與探討預測模型中的參數與結果。
There are a large number of papers and studies about the prediction of stock market which show us the importance role it plays whether for industry or academia. In this article, deep learning models for stock prediction are composed of the architecture of deep learning neural network, long short-term memory, embedding layer and the thought of dropout neural network architecture. Data about 544 stocks listed company at stock exchange market and over-the-counter market are offered to train our models. There are ten features, including day's range (high, low), day’s open, day’s close, day's amount, the balance of margin loan, the balance of stock loan, and net buy and net sell of foreign investment institution, investment trust, dealer. The model will propose stocks making a profit higher than 5% and 3% lower loss of risk which can give some reference for the development of economy and investment. For deep learning, long short-term memory may not be the only choice, that’s why we build another model with convolution neural network and discuss the advantage and disadvantage of the two models.