上市公司的现金分红政策一直以来都是国内外的研究热点，也受到我国监管层的较大关注。信号传递理论表明，现金分红政策可以向市场传递有关公司经营情况的有效信息，投资者从中接收信号并进行买入卖出操作，引起个股收益率显著波动，被称为现金分红的市场反应。然而，长期以来我国上市公司的现金分红具有不稳定、不持续的特点，导致投资者无法从中获取有效信息。自我国证监会颁布一系列半强制分红政策以来，上市公司的现金分红意愿和力度有所提高，稳定性有所改善。因此，有必要对半强制分红政策实施以来，上市公司现金分红的市场反应展开研究，验证其信号传递效应是否有所改善。拟采取事件研究法，检验不同类型现金分红公告引起的个股异常收益率显著性及差异情况，并采用多元线性回归的方法区分盈利信息和股利信息的市场反应。基于事件研究中“相较于现金分红减少或不变，现金分红增加能够显著提高短期个股收益”这一发现，建立现金分红变动的预测模型，以帮助投资者进行买入卖出操作，从而获取超额收益或减少损失。为此，首先利用T检验、和皮尔逊卡方检验筛选特征变量，并分析单一变量与现金分红变动的关系。随后，利用逻辑回归、决策树、随机森林、GBDT、XGBoost算法建立预测模型，并基于特征对样本不纯度减少的贡献，计算特征变量重要性并排序，揭示各因素对现金分红变动的解释力和预测性。 实证研究结果表明，每股现金股利的未预期增加对个股累计异常收益率有显著为正的贡献，且其显著异常收益率集中于公告日前出现，市场上存在信息泄露。预测结果表明，相较于逻辑回归和决策树算法，集成学习算法GBDT、XGBoost能够显著提升对每股现金股利的预测精准率，模型对前10/30/50/100个正类样本的预测精准率达100%/90%/88%/86%。在研究选取的特征变量中，每股盈余同比、留存收益同比、可持续增长率对每股现金股利变动的预测性最强。研究结论支持Lintner股利理论和股利政策的生命周期理论，表明本期利润变动是上市公司做出股利变动决策时考虑的重要因素，相较于衰退期和成长初期的企业，成长期向成熟期过渡阶段、成熟期的企业更有可能增加每股现金股利。

The dividend policy of listed companies has always been a hot topic of academic research in China and foreign countries. It has also has aroused extensive attention of the Chinese regulators. Signaling theory shows that the cash dividend policy can transmit effective information about the company to the market. Investors receive signals and carry out buying and selling operations, causing abnormal return of stocks. However, for a long time, the cash dividend policies of listed companies in China are unstable, which makes it difficult for investors to obtain effective information. Since the release of semi-mandatory dividend policy by CSRC, the willingness to pay cash dividends and cash dividend per share of listed companies in the A-share market has increased. Therefore, It is necessary to study the market reactions of cash dividend policies to verify whether its signaling effect has improved. Based on the findings of the event study, this paper builds prediction models to predict changes of cash dividend per share. The event study method is proposed to test the significance and difference of abnormal returns caused by different types of cash dividend policies. Multiple linear regression is used to distinguish the market reaction of profit information and cash dividend information. Based on the finding that cash dividend increase can significantly improve the short-term stock returns, the paper build prediction models of cash dividend changes. Firstly, T-test and Pearson's chi-square test were used to eliminate irrelevant variables, and to analyze the relationship between single variable and cash dividend changes. Secondly, logistic regression, decision tree, random forest, GBDT and XGBoost are used to establish prediction models for cash dividend changes. Thirdly, the paper reports the importance index of variables to evaluate the predictability of each variable to cash dividend changes. Empirical results show that market reactions are significant positive to the unexpected increase in cash dividend per share. The market reactions to the increase of cash dividends begins before the announcement day, which indicates that there is information leakage in the market. Among all the methods, The GBDT and XGBoost models has the highest prediction accuracy. The precision for the first 10/30/50/100 positive samples is 100%/90%/88%/86%. It is found that the growth rate of earnings per share, the growth rate of retained earnings and the sustainable growth rate are the most important factors to predict the changes of cash dividend per share. The conclusions of our study support the Lintner’ s dividend policy theory and the life-cycle theory. The paper proves that the profitability increase is one of the most important reasons of cash dividend changes. In addition, companies that are in their slow growth and mature stage are more likely to increase cash dividend than any other companies.