Prediction Based GLM And Financial GLMM

How to accurately predict the financial crisis of listed companies is animportant and meaningful issue, but also a very difficult problem. The reason is theaccuracy of the forecasts not only depends on the statistical model, but also is closelyrelated on the selection of variables. This paper starts from the variable selectionbased on two commonly used models—generalized linear models (GLMs) andgeneralized linear mixed models (GLMMs) to build the financial early-warningmodels of listed companies, as well as study the influence of the model selection andvariable selection on the financial prediction of listed companies.In this paper, we select the financial data from120manufacturing listedcompanies during the year of2006to2012to establish related financial indicators.Based on the preliminary analysis, we found strong correlation between theseindicators. Whether a listed company is included ST by the SFC is the dependentvariable of this study. It is obvious that the dependent variable is a0-1variablefollowing the Bernoulli distribution which is a special form of the binomialdistribution. So we need to use generalized linear models (GLMs) to fit this question.Thus, we use stepwise regression, LASSO and SCAD methods of variable selectionto build3generalized linear models. Based on the number of selected variables, theeconomic significance of the variables, the model fittings and the prediction effects,we believe that the most ideal model is SCAD method of variable selection, LASSOand stepwise regression methods are less effective. The method of LASSO is theworst for the prediction of ST companies amang these three models; the method ofstepwise regression selected two bad variables whose economic significance areinconsistent with the financial theory.Next, according to the selected variables from these three methods, weconstruct generalized linear mixed models (GLMMs). As we all know, listedcompanies of different industries are face with different financial risks on average. Taking this for consideration, we combine New Wealth Industry ClassificationStandard and selected variables of these three methods construct three generalizedlinear mixed models. The results show that the overall effect of the three models isvery close, and the model of SCAD is slightly better. In this paper, by comparisonwith generalized linear mixed models and generalized linear models, we find thefollowing advantages of generalized linear mixed model:(1) the part of randomeffects of the model in line with the characteristics of the development of theindustry may clearly explain the reality condition of each industry;(2) the model ismore concise, the number of variables which fixed effects of GLMMs contain aresignificantly less than the GLMs’s.(3) Compared with GLMs, the predictionaccuracy of GLMMs is improved. In addition, comparing with data mining methods,we found that GLMMs are superior to the methods of data mining, embodied in thesimplicity and practicality.This study shows that variable selection is directly related to the accuracy ofmodel predictions in the financial crisis prediction. The method of SCAD variableselection is significantly better than LASSO and stepwise regression methods. Inaddition, the statistical models are also directly affecting the accuracy of the financialprojections. Compared with GLMs, the prediction accuracy of GLMMs hasimproved to some extent, and the models are more streamlined, more in line witheconomic significance.