結合資料離散化與通道獨立於時間序列資料之趨勢預測

Abstract

一直以來股票市場都是投資人所關注的熱門領域，使用深度學習技術進行時間序列資料的研究也逐漸成為趨勢，由於自注意力的架構在自然語言及圖像處理等領域的任務中有著不錯的表現，因此許多研究也試圖利用此架構來提高時間序列任務的預測準確度。然而，對於高波動、高雜訊、及非線性的金融資料而言，深度學習模型往往需要大量的歷史資料來學習，因此較不利於使用原始的自注意力架構。此外，在深度學習模型中，僅使用連續型的技術指標資料往往難以獲得理想的股價走勢預測結果。因此本論文引入了適合多變量長輸入時間序列資料的模型架構並結合資料離散化的方式，來預測股價的上漲、持平、及下跌。實驗結果顯示，對於各項深度學習常見的衡量指標，都能利用這樣的方法來取得顯著的提升，證明此論文所使用的方法對於預測未來股價走勢的有效性。

The stock market has always been a popular area for investors. Meanwhile, deep learning techniques have become increasingly prevalent in analyzing time series data. Since the strong performance of the self-attention architecture for natural language processing and image processing, numerous studies have leveraged this architecture to enhance the prediction accuracy of time series models. However, for financial data with high volatility, high noise and non-linearity, deep learning models often need a large amount of historical data, which makes the original self-attention architecture less suitable. Furthermore, using only continuous technical indicators in deep learning models can lead to suboptimal predictions of stock price movements. Therefore, this thesis introduces a model architecture specifically designed for multivariate time series data with long input data. The model also incorporates data discretization technique to predict stock price rises, no changes, and falls. The experimental results demonstrate that this approach significantly improves the performance of model prediction based on various common deep learning metrics, proving its effectiveness.