標題: | 基於混合式深度卷積神經網路與稀疏編碼設計智慧型期貨交易策略 A hybrid convolutional neural networks with sparse coding for intelligent future trading strategy design |
作者: | 陳柔帆 陳安斌 黃思皓 Chen, Jou-Fan Chen, An-Pin Huang, Szu-Hao 資訊管理研究所 |
關鍵字: | 深度學習;機器學習;智慧交易;交易策略;卷積式類神經網路;卷積稀疏編碼;稀疏編碼;Deep learning;machine learning;intelligent trading;trading strategy;convolutional neural networks;convolutional sparse coding;sparse coding |
公開日期: | 2017 |
摘要: | 本研究提出結合卷積式類神經網路與卷積稀疏編碼的混合型深度學習模型,概念為利用稀疏編碼的三項輸出,將善用其特性分別用不同方式搭配於卷積式類神經網路,建置出新穎的混合型深度學習模型,以達到在小樣本資料的學習當中,獲得更好的學習效果,其一為利用稀疏編碼的特徵字典,無監督式的自動建置資料特徵,取代卷積式類神經網路原本的隨機濾波器;其二為利用稀疏矩陣,取代卷積層的輸出,搭配池化抽取出每筆資料的重要特徵,作為學習的判斷依據,其三則為將每筆資料以稀疏編碼重建過後的資料代替,目的為期望利用此去躁的特性,能使模型有更好的判斷結果。 本研究期望透過提出的混合式深度模型,能讓深度學習模型在學習金融投資市場行為時,有更顯著的發揮,而經由實證研究,發現一項觀察出的規則交易策略,透過模型的學習,可使其有更好的準確率及獲利表現,尤其是稀疏編碼特徵字典取代隨機濾波器的模型表現相對穩定與突出。本研究得出以上的結論,證實此混合型深度學習模型,在改善金融市場的交易策略時,有良好的表現進而形成一智慧型交易策略。 This paper proposed a series of hybrid deep learning methods which combine convolutional neural networks (CNN) and convolutional sparse coding to enhance the performance of intelligent financial data analysis. The idea of this hybrid model is to utilize three meta-results of dictionary learning into deep learning framework. It aims to achieve better accuracy of financial data analysis for future trading strategy design. We proposed three novel hybrid models in this thesis as our major contributions. The first one is named as learned filter CNN model (LFCNN), which substitute a feature dictionary for the random filters in CNN framework. A set of initial filters is constructed by the unsupervised learning approach of sparse coding. The concept of the second model is to replace convolution layer with sparse matrix. This model aims to get better learning efficiency and accuracy based on sparse features. The third model uses the reconstructed data from sparse representation as input, and the purpose is to remove noise and make the decision making more robust. This research expects to enhance the predictability of investment behavior through the proposed hybrid deep learning models. The experimental results show that the proposed models can improve the accuracy and profit performance dramatically than the traditional rule-based strategies. The trading strategy derived from LFCNN model is relatively stable and prominent than the other two method proposed and examined in this thesis. In summary, the hybrid deep learning models are proven to have superior performance on improving the trading strategy and developing an intelligent trading strategy. |
URI: | http://etd.lib.nctu.edu.tw/cdrfb3/record/nctu/#GT070453404 http://hdl.handle.net/11536/141695 |
Appears in Collections: | Thesis |