請用此 Handle URI 來引用此文件:
http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/55461
標題: | 基於混合式擴張卷積改善時間卷積網路預測類股趨勢的能力 Improved Temporal Convolutional Network on Stock Trends Prediction Based on Hybrid Dilated Convolution |
作者: | Yu-Lin Zhou 周宇霖 |
指導教授: | 呂育道(Yuh-Dauh Lyuu) |
關鍵字: | 卷積類神經網路,時間卷積網路,擴張卷積,股價預測, Convolutional neural network,Temporal convolutional network,Dilated convolution,Stock price prediction, |
出版年 : | 2020 |
學位: | 碩士 |
摘要: | 股票市場是個變化劇烈且快速的金融市場,外加有大量雜訊的干擾,使得預測股票市場的漲跌是件困難的事情。類神經網路具有強大的函數模擬能力,因此適合利用股票市場所產生的大量非線性資料來學習並預測。相對於時間序列建模常使用的 LSTM,本論文使用時間卷積網路 (TCN) 此一基於 CNN 但被特別設計的適合處理時間序列的類神經網路,將之用於台灣股票市場漲跌的預測。 擴張卷積是 TCN 的關鍵架構,其指數成長的擴張因子讓 TCN 在相同的參數量下,可以讓感受野 (receptive field) 以指數方式成長,使其獲得足夠多的歷史資訊。然而多項研究指出這種指數成長的擴張因子會導致 gridding 的問題,因而損及類神經網路的預測能力。本論文基於混合式擴張卷積提出覆蓋式擴張卷積網路 (C-TCN) 改善此一問題,並與原本的 TCN 一同比較雙方在預測加權報酬指數與化學生技醫療報酬指數漲跌的能力。實驗結果顯示,在兩個標的各六個時間窗口共十二次實驗中,C-TCN 在其中十個實驗勝過或持平 TCN,有達到改善的效果。 Stock market is a very complex financial system. Stock price can rise and fall dramatically in a short period of time. With a high degree of noise, prediction of stock market is a difficult task. Neural network is a powerful data-driven model that can approximate arbitrary continuous function. This makes it a good choice to learn from the enormous non-linear data generated by the stock market. Instead of using LSTM, which is known for being good at predicting time series, we use temporal convolutional network (TCN) on stock market prediction. TCN is an architecture based on convolutional neural network (CNN), and combines the best practice of recent convolutional architectures for sequence modeling. Dilated convolution is a key of TCN, as the exponentially increased dilation factors can expand the receptive field exponentially without requiring extra parameters. This allows TCN to have longer effective history information in an efficient way. However, many researches pointed out that the dilated convolution using exponentially increased dilation factors suffers from gridding effect, which hampers the performance of the neural network. On top of TCN, we propose C-TCN, which is based on hybrid dilated convolution, to address this problem and compared it with the original TCN on the prediction accuracy of TAIEX Total Return Index and Chemical, Biotechnology and Medical Care Total Return Index. We use 6 sliding windows for each target. The results show that C-TCN is better than or equal to TCN in 10 out of the 12 experiments. |
URI: | http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/55461 |
DOI: | 10.6342/NTU202002065 |
全文授權: | 有償授權 |
顯示於系所單位: | 資訊網路與多媒體研究所 |
文件中的檔案:
檔案 | 大小 | 格式 | |
---|---|---|---|
U0001-2907202022214800.pdf 目前未授權公開取用 | 2.03 MB | Adobe PDF |
系統中的文件,除了特別指名其著作權條款之外,均受到著作權保護,並且保留所有的權利。