卷積神經網路預測時間序列能力分析

Analysis of the Predictive Ability of Time Series Using Convolutional Neural Network

本研究試圖將深度學習的概念與技術，應用於金融商品價格走勢的預測並據以建立一個績效良好的交易策略。主要的方法是將金融商品在一段時間的開盤價、最高價、最低價、收盤價等一維時間序列資料進行二維圖像化。傳統研究方法是從一維的測度去研究時間序列走勢，本研究利用人工智慧（AI）的方法將一維時間序列資料轉化成二維的圖像資料。我們將利用在圖像辨識有卓越表現的卷積神經網路（CNN）進行特徵的萃取，進行金融商品未來漲跌的分類，藉此達到預測走勢的效果，進而建構一套可以穩定獲利的交易策略。本研究的實證結果發現：透過將時間序列二維化的方法，交易模型比單純輸入時間序列數值可以學習到更豐富的訊息，因而達到穩定提升交易績效的成果。由此可知資料處理方式具有強大的影響力也證實了人工智慧的重要性，在臺灣開始發展金融科技之際，期以本研究有助於往後的研究者、金融機構和監理機關更進一步研發相關的技術。

Financial development and technology are always closely related, on the whole, finance has gone through the process of electronization and networking. At the present stage finance is in the process of automatization and intelligentization, and in this paper, we try to apply the concepts and techniques of deep learning to the forecast of price trend of financial products. The main concept is that we transform one-dimensional time-series data of opening price, highest price, lowest price, closing price into two-dimensional planes. Most previous researches used one-dimensional measure to study time-series, here we use two-dimensional measure to study time-series. Then, we use the convolution neural network (CNN), which has excellent performance in image recognition to capture features and make the classification of price trend, so as to achieve the effect of forecasting price trend and construct a trading strategy which can obtain reliable result. The empirical results show that deep learning models can learn better by using the method of visualization than simply inputting time series values and the performance is more stable. Therefore, as Taiwan begins to develop FinTech, it is hoped that this paper will help future researchers, financial institutions, and supervision agencies develop related technologies.