基於深度學習演算法之多變量時間序列趨勢預測:以股市分析為例

Abstract

多變量時間序列趨勢預測在許多領域中具有高度的應用價值，例如，在股市應用中，若是能夠建立有效的股票價格趨勢預測模型就能給投資者帶來許多利潤。目前有關股市預測的研究中，大部分使用了文字探勘、技術指標、循序模式等方法來建立預測模型，並只專注於未來某一個時間點上的價格或漲跌預測。有鑑於此，在本論文中，我們提出了一個多變量時間序列分類機制並將其運用於股市價格趨勢預測，此機制結合以特徵為基礎的分類方法以及以非特徵為基礎的分類方法，並將股市價格之漲跌趨勢預測轉換為多變量時間序列分類問題。我們利用技術指標建構基本特徵，並結合以非特徵為基礎的深度學習方法來自動建構特徵及預測模型。經由一系列真實資料集的實驗評估，整體來說本研究所提出之方法比傳統分類器在預測準確度上具有更好的表現，可以顯現出我們的模型可以有效地預測股票之漲跌趨勢。

Multivariate time series trend prediction is highly valuable in many areas. For instance, in stock markets, if we can build an effective stock price trend prediction model, it will bring more profits to investors. Most existing related researches focused on constructing features and models from text mining, sequential pattern mining, and technical indexes while their target is rise or fall and price forecast at some point in the future. In this thesis, we propose an effective multivariate time series classification framework and apply it on forecasting stock price trends. The proposed framework combines the feature-based methods and the non-feature based method. We used the technology indexes to construct features and combined with the non-feature-based deep learning method to automatically construct the features and the prediction model. A series of experiments were conducted on real data. The method proposed by this work performs better than traditional classifiers in accuracy. It shows that our model can predict stock price trends more effectively.