Research On The Dynamic Modeling Of Business Financial Disreess Prediction

Nowadays, the company's operating environment presents the characteristics of global economic integration and customer demand diversification, and the industry competition is intensifying. It is the uncertainty in the external environment and the internal understanding limitations that make the enterprises face increasing internal and external risks. Financial distress is the most comprehensive and significant representation after the deterioration of the internal and external risk factors. As the stock market and the bankruptcy law become more perfect, companies which falls into financial distress is easier to be delisted or bankruptcy. Enterprises are important capital market participants, and its development has been the concern of many stakeholders, which even can determine the investors'anticipation of the economic trend. Therefore, in the background of global financial crisis, how to effectively predict financial distress has important reality significance.However, the current financial distress prediction research is still concentrated on the selection of single prediction methods based on static data set, or on the simple combination of multiple prediction methods. The combination prediction method for financial distress is not thorough enough, and its main expression is the basic structure of the combination system is set beforehand, which leads a randomness of the model performance, a poor generalization, and the lacking of data-driven dynamic structure selection mechanism. Furthermore, the present method of financial distress prediction is still in the static modeling stage, sample data collected in a specific period of expansion, but corporate financial data will incrementally change by the time passage. So breaking the restrictions of static modeling, building the model which can be dynamically updated by the time goes on, and meeting the new economic environment or business dynamic operating environment requirements, have important theoretical significance.This paper firstly analyzes the meaning of business financial distress, which leads to the internal and external reasons that makes business fall into it. Based on basic theory knowledge of economic early warning, we demonstrate the rationality to use financial ratios as signs of financial distress. Then we put forward the dynamic modeling framework of financial distress, which consists of two angles: first, dynamic selection of structure of multiple classifier system; second, dynamic financial distress prediction modeling for sample data stream. Verifying the theoretical model by data experiments, so the first job is sample selection and data preprocessing. We select 321 pairs of financial distress sample companies and their matching samples from Shanghai and Shenzhen Stock Exchange as experiment samples for the following static data set and dynamic data stream. The initial financial ratios system contains 40 financial indexs, and then two cross-section data is obtained to establish the financial distress model. Through the normal test and mean non-parametric test of financial ratios, it indicates that the vast majority of financial ratios are significantly different between the two types of samples. We apply multiple feature selection method on the two cross-section data sets to acquire multiple feature financial ratio subsets, aiming to reflect the financial distress information content from various angles, and use them as the input of combination model of financial distress prediction in the following sections.For static dada sets, we proposed multiple classifiers selective ensemble financial distress prediction method. First of all, from the perspective of qualitative analysis and quantitative derivation, it demonstrates that combination of multiple classifiers could bring potential benefits and the corresponding condition, so that to explain the need for selective ensemble. A dynamic selective ensemble method based on greedy search and pruning is brought forward. It can dynamically minging combination sytem structure from the baisic classifiers library by acquiring multiple local optimal combination systems. Experiments results shows that: this method is significantly better than the static ensemble system and the individual best single classifier, and this fusion sequence is better than integration in descending order, integration in ascending order and random integration. Then we design a selective ensemble method which is based on genetic search. It uses genes to represent the basic classifiers, so that the combination prediction system can be expressed as a chromosome. Set the predictive accuracy of the combination system as its global optimization objective, it searchs the optimal combination system by genetic operations. Experiments results indicate that it can further improve the accuracy.For the characteristics of sample data incremental flowing in financial distress prediction, combined with the theory of concept drift in pattern recognition, we present a dynamic modeling method of financial distress prediction that can deal with concept drift. Based on deep analysis of the meaning of financial distress prediction, two window size selection mechanisms, namely fixed size window and adaptive size window are designed. It is show in the experiment that adaptive size window can dynamically adjust window size according to the feature of current financial distress concept, its predictive performance is better than fixed size window method in terms of theory and reality. Based on the above, the dynamicselection method of multiple classifiers is combined with the window size adaptive method. It defines the concept of local prediction space of future sample that will be predicted, produces basic classifiers by the bagging method, and dynamically selects one or a few basic classifiers for one future sample based on the performance metric of the local prediction space. The experiment results show that: in most cases, the dymamic selection and integration can further enhance the predictive performance based on the adaptive window.This paper carries out the research of dynamic modeling method for financial distress prediction. It respectively builds models with the static data set and the dynamic data streams. So it can help to enrich the relevant study content of financial distress prediction, and promote the research of financial distress prediction from the static modeling stage to a dynamic modeling stage.