整合資料探勘與類神經網路技術於台灣股市投資決策-以台灣積體電路產業鏈為例

Abstract

在股票投資過程中，投資人如何決定最佳之買賣時點是個非常重要的問題，因此投資人極需一套適當的工具，以協助投資者進行買賣的決策。一般常使用統計方法和人工智慧演算法來為投資者建立決策支援系統，而常被使用的方法有、資料探勘、類神經網路等方法。雖然類神經網路在解決非線性問題以及在財務金融領域上都有不錯的處理能力，但其運作過程中，有如黑箱作業的內部運算不易令人清楚的了解運作過程，而且更難以解釋其原理。而單獨使用資料探勘來對歷史的交易記錄中發掘股市變化的漲跌，也只能找出概略的方向，並無法很精確的去學習影響股市的因素。為彌補彼此的缺點，因此本研究題出先導入關聯法則來補足類神經網路在解釋能力上的不足的缺點，借由資料探勘技術來分析半導體產業中，其股票是否存在漲跌順序的規則，依分析結果找出連動的關係，詳盡的瞭解彼此之間的漲跌情形，找尋出規則後再利用類神經網路模仿人類的思維模式並分析變數與準確度之間的相關性，以建立具有解釋力之理論模型。而能從理論模型中達到學習、預測的目的。 本研究以台積電供應鏈關係所組成的上、中、下的三家公司股票為研究實例，來驗證此整合模式的有效性，實驗結果顯示，本研究所提出整合資料探勘與類神經模式的預測結果皆優於資料探勘的單一預測模式。

How investors determine the best timings to buy and sell stocks is a very important issue in stock investment. Therefore, investors must have a suitable tool to help them to make decisions. In general, people use statistical methods and artificial intelligence algorithms to establish decision support systems for the investors, including data mining and neural network. Although neural network had good effects on solving non-lining problems and financial area, it is hard to explain its internal algorithm such as a black box. But if we only use data mining to discover the ups and downs in the historical dealing transactions, we can only find a rough direction not a precise factor how to affect the stock. For making up the disadvantages, this study present to use association rule to explain what neural network can’t explain. Then, by using data mining technique, we analyze the stock prices of Semiconductor industry to have association rules about rising and falling. Furthermore, we find the correlation from the result and understand the condition of fluctuation. Finally, we find the rules and make use of neural network to simulate the way of thinking of the human and analyze the correlation between the variables and accuracy. From the above, we establish a technical model with explanations and learn to predict from this model. To verify the availability of this model, we use the stock data of supply chain. The experiment results show that the proposed model constructed by the data mining and the neural network is better than the data mining model in prediction.