應用類神經網路於股災期間之股價預測—以台灣成分證券ETF為例

台灣股票市場容易受國際股市影響，股市甚麼時候會崩盤則是各位投資者在意的事情，期望藉由觀察歷史上四次代表性股災，結合類神經網路的運算，預測往後股價的走勢。選取的四次股災分別是：2008年的金融海嘯、2011年的美債危機導致全球股市重挫、2018年的美中貿易最高峰和2020年的新冠肺炎疫情爆發帶來全球經濟衰退疑慮。在這些事件中，一國股價之漲跌容易對其他國家有連鎖效應，特別是股價突然暴跌時連動性相當明顯。因此，本研究將各國主要股市指數、代表原物料物價的期貨、消息面指標和總體經濟分析指標納入變數中，並加以預測指數股票型基金(Exchange Traded Fund, ETF)之價格。本研究應用倒傳遞類神經網路整合台灣發行量加權股價指數、美國紐約道瓊工業平均數、美國紐約史坦普爾500股價指數、香港恆生指數、中國上海綜合股價指數、德國DAX指數、韓國綜合指數、日本日經指數、法國CAC指數、英國FTSE100指數、加拿大多倫多綜合指數、新加坡富時海峽指數、泰國曼谷SET指數、CRB指數、恐慌指數、台灣領先指標綜合指數、景氣對策訊號、海關出口值、機械及電機設備進口值預測近十年殖利率排行前十名的國內成分證券指數股票型基金。本研究採用皮爾森相關係數分析四次代表性股災台灣股市收盤價與各國指標的相關性，數據結果顯示國際之間股價的相關性高。本研究數據結果平均絕對百分比誤差(Mean Absolute Percentage Error, MAPE)皆小於10%，所以，本研究確認在四次代表性股災時間，倒傳遞類神經網路學習國際股價趨勢來預測台灣股票市場是有效的。關鍵詞：倒傳遞類神經網路、多層感知器、價格預測、指數股票型基金、平均絕對百分比誤差

關鍵字

倒傳遞類神經網路；多層感知器；價格預測；指數股票型基金；平均絕對百分比誤差

並列摘要

Taiwan Stock Market is easily been affected by international stock markets. Therefore, the investor pay attention to be alert when the stock market will be tumble-down. With the observation of the historical representative stock market crashes and the calculation by using the artificial neural network, the goal of this research is to predict the future trend of stock prices. This research collected stock price data of four historical stock crashes as the subject. The first stock crash was a financial crisis in 2008. Second, the U.S. debt ceiling crisis occurred in 2011 that caused global stock markets to plunge. Third, the peak of China–United States trade war occurred in 2018. Finally, the outbreak of the COVID-19 occurred in 2020, which brought concerns about global economic recession. In these crashes, the rise and fall of a country's stock price is likely to have a knock-on effect on other countries, especially when the stock price suddenly plummets, the linkage is obvious. Therefore, this study incorporates major stock market indices in various countries, futures representing the prices of raw materials, news indicators, and total economic analysis index into the variables in order to predict the price of Taiwanese Exchange Traded Funds (ETFs). This research uses the back propagation neural network (BPN) to integrate Taiwan capitalization weighted stock index, Dow Jones industrial average in New York, the US New York Standard Poor's 500 stock price index, the Hong Kong Hang Seng index, China's Shanghai Composite Stock Index, the German Der Dax (DAX) Index, Korea composite stock price index, Japan Nikkei index, French Cotation Assistée en Continu (CAC) index, UK Financial Times Stock Exchange 100 (FTSE100) index, Canada Toronto composite index, Singapore Straits Times Index (STI), Stock Exchange of Thailand (SET) index, Commodity Research Bureau Futures Price (CRB) Index, Volatility Index (VIX), Taiwan composite leading index, monitoring indicators, customs export value, machinery and electrical equipment import value forecasted the top ten domestic constituent securities index stock funds in the past ten years. This study uses Pearson's correlation coefficient to analyze the correlation between the closing price of the Taiwan stock market in four representative stock market crashes and the indicators of various countries. The results show that the correlation between international stock prices is high. The mean absolute percentage error (MAPE) of the results is less than 10%. Therefore, this research confirms that, in the four representative stock market crashes, the BPN network is effective to predict the Taiwan stock market by learning international stock price trends. Keywords: back propagation neural network, multilayer perceptron, price prediction, exchange traded fund, mean absolute percentage error