目前股票分析之實證研究,多以技術面或基本面等資訊,對動態的股票市場做預測與研究,但因大部分皆屬落後之資訊,不但常會出現對於市場趨勢的過度反應或反應不足,導致預測準確率降低及無謂的投資損失,且近年來政府大力提倡自由化與國際化,更使機構投資人於台股占有舉足輕重之地位。故本文以倒傳遞類神經網路來探討機構投資人交易行為對台灣股市的影響能力,並與灰預測之股價預測能力作一比較,試圖找出較佳投資準則,以作為投資人之參考依據。 在類神經網路架構下,針對三種技術指標KD、MACD、MA做為輸入變數,並配合「量先價而行」之理論,以三大法人持股週轉率為輸入變數,測試對股價預測能力是否有所提升;在灰預測部分,則採用傳統的GM(1,1)模型,並且利用滾動建模的方式來預測收盤價。 實證結果發現,在預測隔天收盤價的表現,灰預測優於傳統倒傳遞類神經網路,法人持股比率的高低對於倒傳遞類神經網路之預測能力有明顯差異。如再結合灰預測對於股價趨勢之優良預測能力於倒傳遞類神經網路後,改良式類神經網路預測能力有顯著的提升。
Nowadays, the empirical studies are generally using technical or fundamental information to produce a prediction and to analyze the dynamic market. However, the majority of information is not updated efficiently and it often has an over-reaction or under-reaction to the market trend. It is also resulted in lower prediction accuracy and unnecessary investment losses. Moreover, the government sector is strongly advocated economic liberalization and internationalization which makes institutional investors play a decisive role in the Taiwan stock market. Consequently, this thesis uses the Neural Network to analyze the forecasting power of institutional indicators in the stock price tendency. After that, the forecasting power complied with the result of Grey Prediction will be compared and an enhanced investment principle will be recommended in this study. In this study, three kinds of technical indicators will be used, KD, MACD, and MA, as input variables in Neutral Network model. In addition, based on the theory of “volume is the lead indicator of stock price”, the stock turnover of institution’s holding is also used as another input variable to the model. The closed price will be forecasted by using traditional GM (1, 1) model for Grey Prediction section. As the result from the empirical studies, Grey Prediction is better than traditional back-propagation neural network in the next day’s closed price forecasting. That is, the ownership ratio of institutional investors has significant differences on the forecasting power from the traditional back-propagation neural network prediction. Furthermore, the forecasting power could be more effective if the advantages of Grey Prediction and revised neural network model are combined with each other.