The Application Of Logistic Regression With Penalization In Financial Distress Prediction Of Listed Enterprises

As an important subject of corporate risk management,the early warning of financial distress could help to effectively identify and reduce potential financial risks in advance.In this study,financial distress prediction model using reasonable statistical methods based on the financial indicators of listed companies has been established to predict the probability of financial crisis in advance.With reference to the existing literature of financial distress prediction,we choose 41 financial indicators as predictors of financial distress prediction model base on the principles of comprehensiveness,measurability and representativeness.Those predictors that we chose reflect the listed company’s solvency,cash flow capacity,operating capacity,profitability,growth capacity and per share index.In general,there is strong correlation between predictors involved in the financial distress prediction model.But very few studies will consider the correlation structure between predictors.Variable selection methods with penalty could select significant predictors,which leads to a higher prediction accuracy.Inspired by this,variable selection methods with correlated-structure penalty have been adopted to Logistic regression model to construct financial distress prediction model in this study.Firstly,we introduce the application of some useful variable selection methods with penalty in Logistic regression,such as Lasso,Elastic Net(Enet).Secondly,the Weighted Fused Elastic Net(WFEnet)method is extended to the Logistic regression model and derive a detailed coefficient estimation process is derived using Coordinate Descent Algorithm.Inspired by WFEnet method and Enet method,a new penalized variable selection method called the Weighted Fused Mnet(WFMnet)based on the combination of the Enet penalty and correlated-structure penalty has been proposed and extended to Logistic regression.Then,we use Coordinate Descent Algorithm to estimate the coefficients of the WFMnet-Logistic regression model.The accuracy of these methods has been compared with the traditional variable selection methods using simulation.The simulation shows that,the Weighted Fused methods have higher accuracy when there is certain correlation between the predictors comparing to Logistic or Enet method,and the WFMnet method is more stable.Finally,we construct the financial distress prediction model using Logistic regression model combined correlated-structure penalty and Enet penalty.Empirical result shows that,WFMnetLogistic model performs well in terms of variable selection efficiency and prediction accuracy.The WFMnet-Logistic model selects 14 significant predictors,such as equity ratio,asset-liability ratio,working capital ratio.Among them,the regression coefficients of positively correlated variables tend to be similar,and the regression coefficients of negatively correlated variables tend to be different.The WFMnetLogistic model has a strong interpretability.