Research On The Modeling Of Listed Companies Financial Distress Ensemble Prediction

Financial distress prediction is an important issue in corporate financial management and investment decision-making, and its essence is to forecast future financial position and classification, which has always been of great concern in theoretical and practical research. The financial distress will affect the economic interests of the managers, creditors, investors and other stakeholders, and even lead to the national capital market volatility and losses. Under the background of the current global capital environment continues to slump and the financial distress still exist, how to monitor the financial risk factors more effectively and carry out the financial distress prediction has important practical significance.Because of the research about financial distress prediction is late in China, and the internal and external environments of the enterprise are complex and variable, foreign financial distress prediction model is difficult to adapt to the reality of our country. Therefore, there is an urgent need to explore a reasonable financial distress prediction model for the relevant stakeholders to provide policy advice and technical support. However, the majority of the current financial distress prediction model is still concentrated in a single model for prediction, or a combination of several models and simple application of classifier ensemble. The models do not consider the adverse effects of the concept drift of financial data. Given the inadequate of the existing financial distress prediction model, we conduct targeted research on listed companies. On the one hand, the classifier ensemble and its improved method have been studied for the modeling process of the financial distress prediction. On the other hand, dynamic incremental model for financial distress prediction is built, considering the concept drift of corporate financial data.Firstly, we research on the classifier ensemble. The concepts and principles of the classifier ensemble are explained in this paper. Then the key steps of classifier ensemble technology are analyzed, mainly including the generation of the base classifiers, and the selection of classifiers and classifiers output. Three commonly used classifier ensemble algorithm are discussed and compared. For the main problem currently exists in classifier ensemble applied research, the applicability of the classifier ensemble in financial distress prediction are analyzed in detail from group decision making, sample size, convenience, etc.Secondly, we research on the model for financial distress prediction based on the classifier ensemble. For the exsiting models are only simple application by classifier ensemble, we analyze the generalization capability of the classifier ensemble system, and use the diversity of base classifiers and individual classifier performances to promote the generalization capability of the ensemble system. On one hand, RS-Bagging classifier ensemble method was proposed by increasing the diversity of base classifiers. The essence of this method is a double disturbance strategy through disturbance of the sample and the feature space. In this way classifier can keep more difference in ensemble system. Model experiments show that, RS-Bagging ensemble prediction model can make up for the deficiency of the single classifier model. RS-Bagging ensemble model outperforms the original Bagging ensemble prediction model. On the other hand, MTS-Bagging classifier ensemble method is put forward by increasing the individual base classifier performances. Mahalanobis-Taguchi System is applied to Bagging algorithm in the method. Proper features are selected by using Mahalanobis-Taguchi System. In this way, individual classifier performances can be improved. Model experiments show that the MTS-Bagging model suitable for financial distress prediction. The accuracy of the MTS-Bagging model outperforms the original Bagging ensemble prediction model. In addition, for the financial distress prediction, the risk cost of the two type errors is different. Considering the risk cost of misclassification, the error rate of MTS-Bagging model is lower than other modles in the research.Thirdly, we research on the model for financial distress prediction based on sliding time window technology. Because of the data samples of financial distress prediction come from different years, there may be a concept drift problem in the modeling. The concept drift problem is solved by the sliding time window and Mahalanobis-Taguchi System, including feature dynamic selection and sample dynamic selection. The prediction model experiment is carried on the different width of the sliding time window. The results show that the concept drift of the financial data does exist. Select the appropriate width of the sliding time window can reduce the adverse effects of the drift concept.Finally, we research on the model for financial distress prediction based on the incremental learning system. Given the lack of incremental learning ability of existing prediction model, an incremental learning system based on the classifier ensemble is designed. Financial category oriented knowledge is applied to realize the dynamic selection for classifier ensemble subsystem. Then the financial distress prediction model with incremental learning ability is build. Model experiments show that, the financial distress prediction model based on the incremental learning system is stable and adaptable. Incremental prediction model outperforms other models, which is an effective financial distress prediction method.