預測股票趨勢對於市場參與者是一個重要的問題，因為即使預測準確性的微小改進也可能導致比其他人有更好的交易決策。且在深度學習技術的發展下，有愈來愈多研究將結構化股價資訊或二維圖像運用於股價預測中，過去的研究將股價資訊轉換成圖像的方式有很多種，包括直接將一維結構化股價資訊直接轉換(reshape)成二維結構化圖像，或是使用格拉姆角場(Gramian angular field, GAF)技術將一維時間序列轉換成二維圖像，又或者是股價K線圖，但大部分研究僅使用其中一個方法，鮮少有針對同一資料集使用三種圖像表現方式進行分析比較，也較少針對類別不平衡(Class Imbalanced)及時間序列型資料增強(Data Augmentation)這兩個議題進行探討；因此在本研究中，會將台灣股市的歷史股價資訊轉換成三種圖像資料，並改善類別不平衡及使用時間序列型的資料增強方法，實驗結果證實使用上述兩種方法後，在三個圖像表現方式上皆能獲得改善，且K線圖和GAF圖像比二維結構化圖像有更好的分類表現。除此之外，本研究也進一步採用多頭CNN(Multi-head CNN)架構，將GAF圖像輸入進多頭CNN模型，實驗結果證實在台灣股市資料集上多頭CNN比CNN有更好的預測效果。;Predicting stock trends is an important issue for market participants, as even slight improvements in prediction accuracy can lead to better trading decisions than others. With the development of deep learning techniques, an increasing number of studies are applying structured stock price information or two-dimensional images to stock price predictions. In the past, there have been various methods to transform stock price information into images, including directly converting one-dimensional structured stock price information into two-dimensional structured images, using Gramian Angular Field (GAF) techniques to convert one-dimensional time series into two-dimensional images, or stock price candlestick charts. However, most researchs only use one of these methods, and there is little analysis comparing the three image representation methods on the same dataset. There is also less discussion on the issues of class imbalance and data augmentation for time series data.
In this study, historical stock price information from the Taiwan stock market is converted into three types of image data, addressing class imbalance and using time-series data augmentation methods. Experimental results show that both methods improve performance for all three image representations, with candlestick charts and GAF images outperforming two-dimensional structured images in classification performance. In addition, this study further adopts a multi-head CNN (Multi-head Convolutional Neural Network) architecture, inputting GAF images into the multi-head CNN model. Experimental results confirm that multi-head CNN has better prediction performance than CNN on the Taiwan stock market dataset.
