| 摘要: | 本研究運用資料探勘技術建構股票股利行情之預測模型來預測台灣上市公司股利公告後的股票行情，以台灣證卷市場上市公司作為研究對象，研究變數包含股利政策、基本面指標、技術面指標、籌碼面指標，探討個股股東會公告股利政策後，是否有上漲10%之行情，並找出影響股利行情之關鍵因子。研究之目標變數為「公告股利政策後是否有10%上漲行情」，將每一筆資料標註 Y 或是 N 做為識別。本研究以2010年至2021年作為研究區間，使用多種技術指標和財務指標作為關鍵因子，包括股利政策、基本面、技術面和籌碼面等方面的指標，並採用資料探勘技術及特徵選取的方法，識別出影響除權息前股利發佈行情之關鍵因子，以提供投資者在每年度除權息期間作為投資參考的依據。
本研究使用決策樹、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。並透過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。實驗設計使用四種不同的預測資料集，以探討各資料集對預測結果的影響。實驗1包含所有上市公司的資料以及所有相關變數(股利政策、基本面、技術面、籌碼面)，用以評估五種監督式學習演算法中哪一種具有較高的準確性；實驗2選取殖利率高於5%以上的上市公司及殖利率低於5%的上市公司資料，並採用所有變數作為預測資料集，比較殖利率高低是否影響結果，以評估針對高殖利率股票是否能獲得更佳的預測結果；實驗3則包含四個子實驗：(1)使用所有上市公司資料，並採用基本面搭配股利政策的指標作為研究變數；(2)使用所有上市公司資料，僅採用技術面財務指標作為研究變數；(3)使用所有上市公司資料，僅採用籌碼面指標作為研究變數；(4)使用所有上市公司資料，並結合股利政策與籌碼面指標；(5)使用所有上市公司資料，並結合股利政策、基本面與籌碼面指標。這些子實驗旨在評估不同變數組合下哪一組能獲得較高的預測分數。
研究建立的67個研究變數，經特徵選取結果顯示，股利行情關鍵因子為股東會至除權息的間隔天數、殖利率、差離狀態(DIF) 、毛利成長率、每股盈餘增減、震盪量指標、乖離率、投信持股率、自營持股率，而基本面指標搭配股利政策的預測能力優於技術指標的及籌碼面指標。
;This study applies data mining techniques to construct a predictive model for stock dividend market trends, aiming to predict the stock market trends after Taiwanese listed companies announce their dividends. The research targets are listed companies in Taiwan′s securities market. The research variables include dividend policy, fundamental indicators, technical indicators, and capital flow indicators. Examine whether there′s a 10% increase in stock price after the announcement of dividend policy in the shareholders′ meeting of individual stocks, and identify the key factors affecting the dividend market trends. The target variable of the study is "Whether there is a 10% increase in stock price after the announcement of dividend policy", with each piece of data labeled as ′Y′ or ′N′ for identification. This study covers the period from 2010 to 2021 and uses various technical indicators and financial indicators as key factors, including dividend policy, fundamental, technical, and capital flow indicators. Data mining techniques and feature selection methods are used to identify the key factors affecting the market trends prior to the ex-dividend date, providing investors with a reference for investing during the annual ex-dividend period.
This study uses five supervised learning algorithms to construct a predictive model: decision trees, random forests, naive Bayes, support vector machines, and artificial neural networks. The predictive model is trained through ten-fold cross-validation, and the prediction accuracy of the model is evaluated using a confusion matrix. The experimental design uses four different prediction datasets to investigate the impact of each dataset on the prediction results.
Out of the 67 research variables established in this study, feature selection results show that the key factors for dividend market trends are the number of days between the shareholders′ meeting and the ex-dividend date, the dividend yield, DIF, gross profit growth rate, earnings per share changes, volatility index, deviation rate, investment trust holding ratio, and self-operating holding ratio. Furthermore, the predictive ability of fundamental indicators combined with dividend policy surpasses that of technical indicators and capital flow indicators. |
