以循環生成對抗網路預測股價量能動態關係

Abstract

預測股價漲跌幅是投資人需要的資訊。技術分析方法種類眾多，當中量價關係為其主流方法之一。本研究使用2007年至 2017年 台積電股票資料，運用首次提出的方法來觀察量價關係，透過循環生成對抗網路(Cycle GAN)結合卷積網路(Convolution Neural Network)與殘差網路(Residual Neural Network)學習量價間的聯合(joint)作用，並得出潛在量價。並參考系統動態學(system dynamic)，將潛在量價與目前量價的位能差距作為影響質量的勢能，透過神經網路轉換成推力， 市值、稅金、昨日漲跌幅視為質量、摩擦力、彈簧力，模擬現實股價上漲下跌的因素。為了方便使用者使用，本研究再將預測股價結合布林帶(Bollinger Bands，BBands)，透過其延伸指標%b指標(Percent b)決定交易信號。經研究證實，結合股價預測的布林帶，比僅有布林帶的平均投資報酬率提升了30%左右。

The increasing demands for stock market prediction plays an important role in nowadays. Via Price -Volume Relation, we are able to access more insight for investors. We propose a novel generative model based on cyclic-consistent generative adversarial network (CycleGAN), which consists of Convolution Neural Network(CNN) and Residual Network(RESNET) for stock market prediction, in particular, Price-Volume Relation. In the reference to system dynamic, we can simulate potential fluctuation of current stock market via Neural network. Considering all the factors of stock market fluctuation e.g. value of stocks, tax, and current price changes as massive, friction, and spring force in system dynamic, our model is able to simulate the fluctuation and gain the ideal forecast. Our research is based on 2007-2018 TSMC stock dataset. For user friendly purpose, we compose BBands (Bollinger Bands) and stock price prediction model to use its derived indicator , %b indicator, to make trade signal. With BBands and our stock price prediction model, its average ROI(Return on investment) has increased 30% efficiency, which is the better result with merely BBands base on our experiment.