L1-regularized LOGISTIC Regression In The Early-Warning Of Listed Company's Financial Risk

The research of company's financial distress prediction has become a widespread concerned issue by scholars both at home and abroad .Not only has high academic value, but also it has great value of application. Constructing the model of pre-warning in financial crisis provide useful information for enterprise managers and stakeholders to do some decision-makings, which has a great significance to predict, reduce the financial crisis, and improve the management level of pre-warning.At present, most of the models of pre-warning in financial crisis focus on the multivariate statistical analysis theories and machine learning methods. The high dimension and redundant information of the variables in enterprise's financial pre-warning indexes system are against the modeling analysis. It's necessary to do some research on variable selection. Many training algorithms now exist for variable selection and model training are mutually independent. This paper uses the regularized method for estimation of pre-warning model in financial crisis. It sets some coefficients to 0, which improves the prediction accuracy and gives an easily interpretable model.The sample data is selected from financial statistics three years before about listed companies titled with ST between 2007 and 2009 as well as the matching sample. First of all, the K-S test and Mann-Whitney U test is used to choose the financial indexes which have a significant difference between the two samples, there are 16, 22 and 24 indexes in year T-3, T-2 and T-1 respectively. Then we use these data to establish a model based on L1-regularized logistic regression for corresponding year. The experimental results show that choosing an optimal penalty parameter for the loss function can not only yield sparse models, but also obtain better predication accuracy (respective 71.43% for T-3, 82.6% for T-2, 94.29% for T-1). Simultaneously, the correlation coefficient matrix indicates that the relativity of output variables is weaker than the input variables of the model.In order to test the validity of financial distress prediction model, we use the same samples and financial data to establish a logistic regression model and a L2 regularized logistic regression, through comparing the accuracy of these three pre-warning models, we find that the L1 regularized logistic regression can obtain a higher accuracy and carry out variable selection at the same time, it also gives an easily interpretable model.