Dynamic Modeling Research Of Corporate Financial Distress Prediction Based On Imbal Anced Time-Series Data Batches

Financial distress prediction research has always been greatly concerned by theoretical and practice circles.Once companies suffer from financial distress, it will not only affect the survival and development of companies itselves, but also affect the interests of other stakeholders, and even trigger the country’s financial crisis.Under the the background of international financial crisis still continuing, weakness of the global economic recovery and the heavy downward pressure of China’s economic, it will have important practical significance that how to build a more scientific and effective corporate financial distress model timely monitoring financial riskIn recent 80 years the domestic and foreign scholars have carried on the thorough research of financial distress prediction theory, and abundant achievements have been presented, however, the current research is still insufficient.On the one hand, most of study is based on balance data set, ignoring the fact that financial distress company accounted for only a few percentage of all the listed companies,which will overestimate the model’s prediction performance.On the other hand, many scholars use the static transverse panel data,without considering the incremental flow characteristics and concept drift problems of the financial data.Models based on static transverse panel data may not effectively predict the new samples within the new concept of financial distress. So in order to establish a more realistic and more effective model, according to the imbalance and timing characteristics of the data, in this paper, based on double perspective of imbalance and concept drift, tries to build dynamic model of corporate financial distress prediction based on imbalanced time-series data batches.The topic of this paper which is dynamic modeling research of corporate financial distress prediction based on imbalanced time-series data batches was abstracted after summarizing the theoretical literature and investigation datas.Then this paper uses the interdisciplinary, qualitative analysis, quantitative analysis and comparative analysis, empirical analysis and other research methods to research the topic from three aspects of theoretical basis, model building empirical research.Firstly, this paper puts forward the meaning and process of the dynamic modeling of financial distress and studies the dynamic modeling theory basis of corporation financial distress, which provide solid theoretical support for the research on the model construction.Secondly, this paper focuses on dynamic modeling research of Financial distress prediction based on unbalanced data and concept drift.In this paper, we use SMOTE sampling and data set segmentation to address the unbalance problem and adopt sliding time window for the financial distress concept drift. The models take SVM, logit and MDA as the base classifiers. For comparing and analyzing the effectiveness of the models in the treatment of imbalance, this paper constructs the financial distress prediction dynamic models respectively based on liding time window and the data set of segmentation; in order to analyze the effectiveness of sliding window in dealing with the problem of concept drift, the paper builds the 2003(T-2),2005(T-2) static model as a comparative models.In the end, the paper collects 373 financial unnormal sample Firms and the 1119 financial normal sample Firms with the time interval of 2003 to 2012 from China listing corperation.In the MATLAB environment, the empirical research is conducted. The experimental results show that the dynamic model of financial distress based on sliding time window is effective, and for SVM and Logit classifier, financial distress prediction dynamic model based on the sliding time window has better performance than financial distress prediction of dynamic model based on the data set of segmentation, for MDA classifier, financial distress prediction dynamic model based on the data set of segmentation has better performance than financial distress prediction of dynamic model based on the sliding time window. In addition, in the empirical research, the SVM is the most sensitive to imbalanced classification problems,while the Logit classifier is lesser and the MDA classifier is hardly sensitive to.