Research On Dynamic Financial Distress Prediction Of Listed Companies Based On Integration Methods

The financial distress can give rise to serious economic loss for enterprise personnel and stakeholders.The good model of financial distress prediction can obtain effective information by mining financial data,which can monitor financial situation constantly,forecast the occurrence of financial distress and provide decision basis for managers.It can play an important role on the healthy development of enterprises.Hence,it is an important matter in the academic community that how to enhance the accuracy for the model of financial distress prediction.At present,the integration models computing static balanced data has been important methods to research financial distress prediction,but these methods construct some simple models commonly,and these models are combined simply or integrated by some particular algorithms,which don't systematically consider conditions to build integration model,and don't consider dynamic selection of classifiers in the process of integration.In addition,the static balanced data still are adopted to predict financial distress in most literatures.The sample data,which are used to predict,are static balanced financial data which are collected in the certain times.It doesn't consider the dynamic characteristic that these financial data can increase gradually with time change,and doesn't consider imbalanced characteristic that,in fact,the number of distressed enterprises is great less than the number of health enterprise s.According to the shortcoming above,the listed companies in our country are considered as the object in this paper,and two types of samples are researched respectively: One type of samples is the static balanced data.The selective integration model ar e constructed under the condition that the integration model needs to be satisfied,which constitutes the combination pattern of dynamic selection of classifiers that has the best prediction result.The other type of samples is dynamic imbalanced data flow.The dynamic prediction model are constructed which can dynamically select samples and deal with imbalanced data.This model can solve the problem of financial data increment,and improve the ability to deal with imbalanced data.First,the sample and the financial indication system are selected.During the process of the selection of samples,648 corporations are selected as samples from the manufacturing industry of the A share market in Shanghai and the Shenzhen Stock Exchanges in the year span of 2007 –2016,based on the theories financial distress and financial distress prediction,and the basis of the sample selection,and the analysis which companies belong to the distressed businesses and which companies belong to the health businesses.The distresse d companies should come from the specially treated companies,and they are not specially treated in the first three years.The health companies should be normal enterprises all the time.The number of samples was selected according to 1:3 of the distressed and health samples.37 indicators are selected based on the principle of financial indicator selection.Financial indicators are removed which are not obvious difference by the normality test and nonparametric test.The sample data are determined finally.Second,the nonlinear manifold learning algorithm is adopted to reduce the data dimensionality.Based on the introduction of the methods of the dimensionality reduction,it analyzes the advantages that the dimensionality of the sample data of financial indicators is reduced by the nonlinear manifold learning algorithm.The theories of ISOMAP ? LLE and LE algorithms are described.These algorithms are used to reduce the dimensionality of financial indicators in order to build multiple feature subsets which ca n reflect the data characteristics of financial information from multiple angles.These data are looked as the basis of the follow-up work.The advantages of nonlinear manifold learning algorithms are proved by experiments.Third,a two-stage selective ensemble model is proposed to predict the financial crisis dynamically for static balance d samples.It analyzes the conditions which the ensemble method needs to satisfy,namely the accuracy and diversity of the base classifier,and discusses the necessity of selective ensemble.According to these conditions,the framework of two-stage selective ensemble model is constructed,and the idea of this framework is illustrated.To enhance the accuracy of classifiers,the kernel function is adopted to improve Fuzzy Self-Organizing Map(FSOM),which can solve the limitation that FSOM demand data to satisfy the spherical distribution.To increase the diversity of the base classifiers,3 kinds of kernel functions are adopted to improve KFS OM to construct 3 models.The three feature subsets,which are calculated by ISOMAP ?LLE and LE,are calculated by 3 models respectively,which can obtain 9 diverse base classifiers.Next,it illustrates how to construct the two-stage selective ensemble model: the first stage,the classifiers are sorted by 3 rules,and the stepwise forward selection method is used to select base classifiers to construct the 3 selective ensemble models.The second stage,3 models are integrated again by different types of companies to construct the final se lective ensemble model.The empirical results showed that the two-stage selective ensemble model can enhance the accuracy of dynamical financial distress prediction when it deals with the static balanced data.Last,the Learn++ model,which is based on the sliding time window and weight sampling,is proposed to predict the financial crisis dynamically for dynamical imbalanced data flow samples.To solve the problems of the concept drift of financial distress and imbalanced data,the incremental learning mod el of Learn++ is improved by sliding time window and weight sampling.Because the data segment is divided subjectively by the Learn++,it is easy to cause the problem of the concept drift for the data in the same training set.The method of sliding time window can solve this problem.The sample weights of the minority class are increased by the weighted sampling method,which can enhance the ability to deal with imbalanced data.To meet the demand of the model on the performance of base classifiers,Probabilistic Neural Network(PNN)is improved by Back-Propagation algorithm(BP)to solve the problems of difficultly estimated parameters and high computational complexity.The empirical results showed that this model can enhance the accuracy of dynamical finan cial distress prediction when it deals with the dynamical imbalanced data flow.The integration method of dynamical financial distress prediction is researched in this paper that,which enriches the idea of built model for the static balanced data and the dynamical imbalanced data,and has the important theoretical and guiding significance.