運用機器學習技術於個股填息預測：以台灣上市公司為例;Applying Machine Learning Techniques to Predict Dividend Recovery: Evidence from Listed Companies in Taiwan

NCUIR > School of Management at National Central University > Executive Master of Information Management > Electronic Thesis & Dissertation > Item 987654321/98197

Please use this identifier to cite or link to this item: https://ir.lib.ncu.edu.tw/handle/987654321/98197

Title:	運用機器學習技術於個股填息預測：以台灣上市公司為例;Applying Machine Learning Techniques to Predict Dividend Recovery: Evidence from Listed Companies in Taiwan
Authors:	張瑋倫;Chang, Wei-Lun
Contributors:	資訊管理學系在職專班
Keywords:	資料探勘;股市預測;除權息;填息天數;填息預測;Data Mining;Stock Market Prediction;Ex-Dividend;Days to Recovery;Ex-Dividend Recovery Prediction
Date:	2025-06-17
Issue Date:	2025-10-17 12:28:39 (UTC+8)
Publisher:	國立中央大學
Abstract:	本研究針對台灣股票市場上市公司，結合機器學習與特徵選取技術，建構預測個股是否能於除息後7天內完成填息之模型。研究資料涵蓋2015年至2019年間共554檔個股，累計2,673筆資料，涵蓋殖利率、股利政策、基本面、技術面、籌碼面及自定義變數等99項指標。研究採用隨機森林、決策樹、貝氏分類器、支援向量機與類神經網路等5種監督式學習演算法進行建模分析，並以AUC、CA、F1、Precision與Recall等指標作為模型評估依據。為強化模型效能，本研究運用特徵選取方法，篩選出對預測填息天數具有高度影響力之關鍵變數。實驗結果顯示，殖利率及多項自定義衍生統計量變數為影響填息天數之關鍵因素，而特徵選取與資料組合策略亦對模型效能有顯著影響。經評估後，隨機森林模型在各項指標上表現最為優異，能有效提升預測準確率。本研究成果不僅有助於學術界深入理解除息後填息現象之影響因子，亦可作為實務界投資人於除息旺季期間，進行短期資金配置與風險管理之參考依據。未來研究建議可擴大樣本範圍、引入多元特徵選取與降維技術，並納入宏觀經濟指標與應用深度學習架構，以提升預測模型之解釋力與實務應用價值。;This study focuses on listed companies in Taiwan′s stock market and integrates machine learning and feature selection techniques to construct a predictive model for determining whether a stock can recover its dividend payout within seven days after the ex-dividend date. Using 2,673 records from 554 stocks between 2015 and 2019, and covering 99 variables including dividend yield, dividend policy, fundamentals, technical indicators, ownership structure, and custom features, five supervised learning algorithms—Random Forest, Decision Tree, Naïve Bayes, Support Vector Machine, and Neural Network—were employed. Model evaluation was based on AUC, Classification Accuracy, F1 score, Precision, and Recall. Feature selection methods were applied to enhance model performance by identifying influential variables, with results indicating that dividend yield and derived statistical features are key factors. The Random Forest model outperformed other algorithms across all metrics, effectively improving prediction accuracy. These findings deepen the academic understanding of post-ex-dividend price recovery and provide practical guidance for investors during dividend seasons. Future research may expand the sample range, incorporate advanced feature selection and dimensionality reduction techniques, and apply macroeconomic variables and deep learning architectures to enhance model interpretability and practical value.
Appears in Collections:	[Executive Master of Information Management] Electronic Thesis & Dissertation

Files in This Item:

File	Description	Size	Format
index.html		0Kb	HTML	33	View/Open

社群 sharing

Loading...