本研究使用台灣證卷市場上市公司填權息資料做為研究，運用資料探勘技術建構股票填息之預測模型，研究變數包含殖利率、股利政策、基本面指標、技術面指標、籌碼面指標，探討個股除權息日後7天內，能否成功完成填權息，並找出填權息的關鍵因子。研究之依變數為「7天內完成填權息否」，將每一筆除權息資料標註Y或是N做為識別，由此欄位分類7天內完成填權息及7天內未完成填權息的資料，選用決策樹、隨機森林、樸素貝葉斯、支援向量機與類神經網路五種監督式學習演算法來建構預測模型。 實驗設計製作三份預測資料集，分別為全部資料集、高殖利率資料集、低殖利率資料集，使用五種演算法，及十折交叉驗證來訓練預測模型，最後使用混亂矩陣來評估模型預測準確度。研究建立的67個研究變數，經特徵選取結果顯示，填權息關鍵因子為殖利率，技術指標的預測能力優於基本面及籌碼指標，技術指標具有效性判斷預測7天內完成填權息否。 本研究另運用特徵選取技術建立資料集，來進行填權息預測，實驗結果在提升預測準確度並未有顯著的幫助，部份演算法出現預測準確度下降的現象，顯示本研究透過特徵選取來降低特徵空間，並無法幫助模型提升預測準確度，但實驗結果顯示研究所使用的研究變項，不存在會造成干擾預測結果的特徵因子，特徵選取功能可以做為評估關鍵因子外，也可以使用來判斷特徵是否會造成模型干擾，或是排除多餘無效的特徵，特徵選取提供不同的驗證方式來探索預測準確度。;This study is to predict the stock price if it can be reverted to previous price within 7 days of the ex-dividend date and find out the key factors for this by using the data of Taiwan listed company to build the prediction model with data mining technology. The variables of prediction model are dividend yield, dividend policy, fundamental index, technical index, chip index. The dependent variable is that the predict the stock price if it can be reverted to previous price within 7 days of the ex-dividend date and mark the data with Y and N to verify with this variable. We use five kinds of supervised machine learning model which includes Decision Tree, Random Forest, Naïve Bayes, Support Vector Machine and Neural Network to build the prediction model. This study builds three prediction datasets which are all dataset, high-yield dataset, and low-yield dataset for training the prediction model by using five kinds of supervised machine learning model and 10 – fold cross validation. We also use the confusion matrix to evaluate the precision of prediction models. There are 67 variables in this study, and we find the key variable is dividend yield by using feature selection and technical index is with good predictability on prediction model than fundamental index, and chip index. Technical index can be more validity on predicting the stock price if it can be reverted to previous price within 7 days of the ex-dividend date. We also use the feature selection to build the dataset, and it doesn’t be with good validity on predicting and there is low precision in some prediction model. So, it can’t be raised the precision by using feature selection. But we don’t find the confounding variable in the selected variables. The feature selection can be used to evaluate the key variables and check if there are any confounding variables.
