如今处于特殊的“双轨制”监管时期，风险分析从业人员与投资者具备的经验参差不齐，本文的模型或方法可以在现有研究的基础上，提供一种具有一定分类与预测能力的补充辅助工具，有利于监管机构与投资者更好地应对上市公司被强制退市的极端风险。此外，结合监管博弈理论，本文延伸性地使用构建的融合模型验证注册制与核准制下不同样本的风险。通过查阅以往有关强制退市风险的研究文献发现，一部分文献是基于资本市场相关理论的定性研究；另一部分文献，主要针对财务风险展开研究。财务舞弊识别是应对强制退市风险的重要组成部分，但是只考虑财务因素不够全面，忽视了风险间的相互作用。因此，本文结合强制退市规定的影响因素与分析方法的可行性，对具体的强制退市规定进行了重分类，从财务状况、公司治理、市场情绪的维度更加全面地考虑了强制退市风险的驱动因素，可以反映金融风险分析理论中常见的财务、合规、市场与流动性风险。本文获取了2011年至2020年的A股上市公司的有关数据，对数据进行样本不平衡处理、数据标准化处理、特征指标降维之后，基于Stacking方法融合了随机森林、支持向量机、BP神经网络算法的基学习器，以Logistic回归作为元学习器，对正负类样本赋予3:1的误分类代价权重作为优化样本不平衡的超参数。本文结合交叉验证、网格搜索等方法构建了强制退市风险模型，相对单一模型的效果与稳定性有一定的提升，可以作为监管者与投资者风险识别与管理的辅助手段。首先，本文使用混淆矩阵及生成的查准率、召回率、F1 Score、准确率评价模型效果，BP神经网络模型的召回率与稳定性不如Stacking融合模型。对比前人研究成果，本文进行了更细致地特征处理，考虑了更全面的风险因素，明显提高了模型的效果与普适性。本文以案例的形式，说明了文中构建的数据挖掘模型可以提前识别出由于信息不对称、串通舞弊、合理化解释等原因而被掩盖的风险。此外，本文通过监管博弈分析注册制对于上市监管带来的积极影响，并将注册制与核准制下不同样本输入融合模型。研究结果显示，注册制监管政策对上市公司管理具有更加积极的作用，对于更加良性的资本市场运作可以起到促进作用。最后，本文提出了后续研究或实际应用过程中可以做出改进的方向。通过增加获取样本数据的频率，更加及时地实现风险识别与管理；根据应用场景与需求的不同，尝试其他改进算法进一步优化模型，并将其模块化地嵌入风险管理系统，有助于提高风险预判与识别的能力。

Now that we are in a special "dual-track" regulatory period, the experience of risk analysis practitioners and investors is uneven. Based on the existing research, the model or method in this paper can provide a supplementary auxiliary tool with certain classification and prediction capabilities, which will help regulators and investors better deal with the extreme risk of listed companies being forced to delist. In addition, combined with the regulatory game theory, this paper uses the constructed fusion model to verify the risks of different samples under the registration system and the approval system.By reviewing the previous research literature on the risk of forced delisting, it is found that part of the literature is qualitative research based on capital market-related theories. The other part of the literature mainly focuses on financial risk research. The identification of financial fraud is an important part of dealing with the risk of forced delisting, but it is not comprehensive enough to only consider financial factors, ignoring the interaction between risks. Therefore, this paper reclassifies the specific mandatory delisting regulations based on the influencing factors of the mandatory delisting regulations and the feasibility of the analysis method. this paper considers the driving factors of forced delisting risk more comprehensively from the dimensions of financial status, corporate governance, and market sentiment, and can reflect the common financial, compliance, market, and liquidity risks in financial risk analysis theories.This paper obtains the relevant data of A-share listed companies from 2011 to 2020. After sample imbalance processing, data standardization processing, and feature index dimensionality reduction are performed on the data, this paper integrates random forest, support vector machine, and BP neural network based on the Stacking method. The basic learner of the network algorithm uses Logistic regression as the meta-learner, and assigns a 3:1 misclassification cost weight to positive and negative samples as a hyperparameter for optimizing sample imbalance. This paper combines cross-validation, grid search and other methods to construct a forced delisting risk model. Compared with a single model, the effect and stability of this model have been improved to a certain extent, and it can be used as an auxiliary means for regulators and investors to identify and manage risks.First, this paper uses the confusion matrix and the generated precision rate, recall rate, F1 Score, and accuracy rate to evaluate the model effect. The recall rate and stability of the BP neural network model are not as good as the Stacking fusion model. Compared with previous research results, this paper has carried out more detailed feature processing, considered more comprehensive risk factors, and significantly improved the effect and universality of the model. In the form of a case, this paper illustrates that the data mining model constructed can identify the risks that are concealed due to information asymmetry, collusion, rationalization and other reasons in advance.In addition, this paper analyzes the positive impact of the registration system on listing supervision through the regulatory game, and inputs different samples under the registration system and the approval system into the fusion model. The research results show that the registration-based regulatory policy has a more positive effect on the management of listed companies and can promote a more benign capital market operation.Finally, this paper proposes directions for improvement that can be made during follow-up research or practical application. By increasing the frequency of obtaining sample data, risk identification and management can be realized in a more timely manner. According to different application scenarios and requirements, people can try other improved algorithms to further optimize the model, and embed it into the risk management system in a modular manner, which will help improve the ability of risk prediction and identification.