應用卷積神經網路辨識股票K線影像改善投資組合策略之研究

Abstract

本研究透過卷積神經網路（CNN）模型進行K線影像預測股票報酬率與成交量，並藉由交乘比較分析以驗證投資組合績效。研究對象為2010年至2017年之台灣上市公司，區分前、後四年為訓練期及測試期進行投資組合實證分析，將股票K線圖之開盤價、最高價、最低價和收盤價等時間序列資料轉換為二維圖像資料，利用深度學習之卷積神經網路的圖形辨識能力，進行特徵分類以預測股價報酬率及成交量。根據模型預測出之報酬率與成交量組別所分類的股票建構投資組合，並以Fama and French（1993）三因子模型及Carhart（1997）四因子模型檢定是否具有顯著超額報酬。實證結果發現前期實際成交量與K線圖形預測報酬率組別所分類的股票建構投資組合可獲得超額報酬，推論以K線影像建構投資組合時必須考慮成交量，將有助於提升投資組合超額報酬的能力。本研究驗證深度學習技術應用於投資策略的可行性，可作為市場參與者投資決策之參考依據。

This study aimed to use a convolutional neural network (CNN) model to predict stock returns and trading volume with a candlestick chart, and cross analysis was used to verify portfolio performance. Companies listed on the Taiwan Stock Exchange at any period from January 2010 to December 2017 were analyzed. Data for the first and second half of the study period were used as training and testing data, respectively. We converted time-series data, including those on the opening price, highest price, lowest price, and closing price of the stock on the candlestick chart, into two-dimensional data. The graph recognition capabilities of a deep learning CNN were used to classify the features of the stock return and predict the trading volume. We classified the portfolio of stocks according to the stock return predicted by the model and the trading volume group and applied three-factor and four-factor models to test for any excess returns. The results demonstrated that the stocks classified by actual trading volume and the candlestick-chart-predicted return rate achieved excess returns. According to the empirical results, trading volume must be considered when constructing an investment portfolio with a candlestick chart to enhance the ability of the portfolio to exceed returns. This study verified the feasibility of applying machine learning to investment strategies and serves as a guide for investors.