| 摘要: | 股票市場的熱絡程度可以代表一個國家經濟的興衰，股價反映的通常都是公司本身的產業前景，股價預測向來都是非常熱門且具挑戰性的研究議題，但由於金融市場的波動容易受到許多外在因素影響，而導致其難以被準確的預測。在深度學習技術發展下，有愈來愈多研究將資訊技術運用於股價預測中，但目前現有的研究都只分析結構化股價資訊，鮮少有針對二維圖像進行分析的研究，也較少針對不同的深度學習技術運用在股價預測上之表現進行比較；因此在本研究中，會將公司的歷史股價資訊轉換成圖像資料，探討在股價預測上運用股市技術圖形，是否會比結構化資料有更卓越的表現。並進一步搭配多種不同的技術指標，探討增加技術圖形中資訊的豐富度和複雜度，是否有助於模型在股市技術圖特徵上的學習。在深度學習模型選擇上，則運用深度學習中之卷積神經網路技術（2D-CNN、VGG16）。此外，混合式深度學習模型和集成式學習在過去許多研究中都已被證實比單一深度學習模型表現來得卓越，但卻沒有研究將混合式深度學習模型運用於股價預測上；因此，本研究將針對混合式深度學習模型（2DCNN-LSTM和VGG16-LSTM）進行探討，比較混合式模型在股價預測上是否能創造出比單一模型更好的表現。
從本研究之實驗結果顯示，運用非結構化圖像資料作為輸入資料，比過去研究運用一維時間序列型資料有更好的預測表現。此外，在實驗中亦證實，增加股市技術圖形資訊量的豐富度，有助於模型在特徵上的學習，提升在股價預測的表現。然而在模型的比較上， 儘管VGG16-LSTM在多數的實驗AUC結果都比VGG16突出，但統計檢定結果顯示此兩種技術之間並無顯著差；而在運算時間上，VGG16-LSTM所花費的時間成本相較於VGG16來得高出不少。因此，若考量到時間成本，在股價預測上選擇單一的深度學習模型，便可得到不錯的效果。
;Stock prediction is one of the most challenging tasks for investors and researchers because the stock market is extremely unstable and volatile due to several factors such as economic, politics, investor sentiment, and more. In the last decade, deep learning techniques start getting more attention, and recent studies have attempted to apply these algorithms to build a model for stock prediction, but most of the studies are focus on using structured data (numerical data), instead of unstructured data (image data). Therefore, in order to understand whether the graph-based technical indicators are the better input data type than traditional structured one, we converted the numerical stock data into stock charts, and combine it with some commonly used technical indicators. In this work, we utilize two CNN-based models to extract features from stock charts. Moreover, we employ one ensemble learning and two hybrid deep learning frameworks, stacking ensemble, VGG16-LSTM, and 2DCNN-LSTM, for performance comparison with the single models.
The results indicate that using the stock charts as the input data for the deep learning models has better performance than numerical data on stock prediction. We also found that the more stock information added to the stock charts, the better performance we can get. Although VGG16-LSTM gets a higher AUC rate than VGG16, the independent sample T-test showed that there is no significant difference between VGG16 and VGG16-LSTM. We dig deeper into the computation time, and we found that VGG16-LSTM takes more training time than VGG16. Therefore, considering the time costs, there is no need to choose VGG16-LSTM in this work. |
