基于高管连锁网络的上市公司财务困境预测研究

Financial Distress Prediction of Listed Company Based on Executive Network

上市公司财务困境的预测问题是金融学领域的热点研究问题之一,科学的财务困境预测有利于企业提高预警管理水平。高管连锁网络通过高管在多家公司兼任的方式在企业中形成网络,使企业之间的决策越来越紧密关联,对企业的行为和绩效产生影响,而现有的财务困境预测模型忽视了高管连锁网络的影响。针对具有网络结构的样本数据,本文提出了带样本网络结构的Logistic模型,根据系数是否受网络结构影响将变量分为网络结构变量和非网络结构变量,不同样本对应的网络结构变量系数是不同的,通过拉普拉斯(Laplacian)二次惩罚函数鼓励样本网络中有连接的样本对应的网络结构变量系数具有相似性。蒙特卡洛模拟表明该模型要优于其他方法。最后,本文应用所提方法基于高管连锁网络进行了财务困境预测研究。结果表明,将高管连锁网络信息纳入财务困境预测模型,可以提高模型预测的准确率。

The prediction of financial distress among listed companies has perennially been a focal point in financial research.The scientific model is conducive to preventing financial distress and improving the early warning management of the crisis.From a methodological perspective,the financial distress prediction problem can be framed as a binary classification issue,where company information acts as explanatory variables for the prediction model.The output is binary,with 1 indicating a company facing financial distress and 0 indicating a company not facing financial distress.Among various financial distress prediction methods,the logistic model is widely used due to its advantages of simple calculation and straightforward coefficient interpretation.Hambrick and Mason(1984)suggested that the psychological factors such as internal cognition,emotions,and values of executives determine their decision-making behavior,thereby significantly impacting business management,financial condition,and future development.With China s rapid economic development,enterprise investment,mergers and acquisitions,and group operations have resulted in an increasing number of directors,supervisors,and senior management personnel concurrently holding positions in two or more enterprises,forming a chain network of executives.The executive chain network embeds the network in the enterprise through the connection of executives,which makes enterprises more and more closely related and has a significant impact on enterprise behavior and performance.Hence,it becomes imperative to integrate the executive chain network into the model when predicting the financial distress of listed companies.However,existing financial distress prediction models have largely overlooked the impact of the executive network.We propose a logistic model that incorporates prior information regarding the sample s network to handle data with a network structure.We categorize variables into structural variables and non-structural variables based on whether their coefficients are influenced by the network structure.The coefficients of structural variables are allowed to vary across different samples.The similarity of the structural variable coefficients corresponding to the connected samples in the sample network is encouraged to be similar by the Laplacian quadratic penalty function.The first part of objective function is the negative log-likelihood function,and the second part is the Laplacian quadratic penalty function.The tuning parameter is selected by five-fold cross-validation.If the tuning parameter is 0,objective function reduces to traditional logistic method.The prediction process involves three steps.Firstly,the model is trained based on the samples and sample networks of the training set,and we get the estimation of the coefficients of non-structural variables and the structural variables corresponding to the training set.Secondly,the coefficient of the structural variables corresponding to the new samples is first calculated based on the sample network.Finally,the explanatory variables and estimated coefficients are utilized for prediction.Section 3 presents simulation studies to evaluate performance of the proposed method.The proposed method is compared with traditional logistic models,neural networks(NNets),random forests(RF),support vector machine models(SVM1 for sigmoids and SVM2 for polynomials),and decision tree models.Monte Carlo simulation results demonstrate that the proposed method performs better than other methods.This suggests that considering the sample network structure can improve the effectiveness of parameter estimation and prediction for new samples.Furthermore,it is evident that in cases where the sample size is small and the variables possess a special structure,conventional black box models exhibit subpar performance.We employ the proposed method to forecast the financial distress of listed companies.To ensure the availability of data,listed companies marked with“*ST”and“ST”are treated as samples of financially distressed companies.Using the data of listed companies from year t 2 to predict whether they will be marked with“*ST”and“ST”in year t.The explanatory variables serve as the foundation of the entire predictive model,and the selection of scientifically sound indicators is paramount.We select 38 indicators,and the specific indicators are detailed in Table 1,with descriptive analyses provided in Table 2.The data used to construct executive network is sourced from CSMAR database.If two listed companies share at least one identical executive,they are deemed connected in the network.The prediction results show that the prediction performance of the proposed method is significantly better than others.Thus,incorporating the executive network into the model can improve accuracy.In order to analyze the coefficient estimation,we established the proposed method and the traditional logistic model based on the entire sample dataset.The estimated values corresponding to these two methods are depicted in Figures 7 and 8.In future research,there is potential to integrate more intricate network structures into financial distress prediction models,while employing variable selection methods to handle high-dimensional data effectively.Additionally,this study primarily focuses on the logistic model with a sample network structure,which could be expanded to include other models such as multi-class logistic regression and Poisson regression.Exploring the application of different models across various fields would be a valuable avenue for further investigation.