Research On Financial Early-warning Model Of Manufacturing Listed Companies Based On Random Forests

With the increasing market competition, the risk of company financial crisis is alsogradually increasing. For managers, investors, creditors and other stakeholders of a company,it is one of the biggest concerns of them whether financial distress can be effectively predicted.Financial early-warning has a history of more than80years, roughly undergoing two stages:statistic analysis and data mining. In recent years, Support Vector Machine (SVM) andArtificial Neural Network (ANN) based on artificial intelligence win a wide range ofapplication in the field of financial early-warning and effectively raise efficiency of theearly-warning. However, because of the unbalancedness of financial early-warning problemand noise as well as the complexity of data distribution, these methods cannot gain satisfyingresults. In addition, such methods do not have enough explanatory ability for the model,which greatly degrades the credibility of the model. Therefore, this paper introduces acombination classification algorithm——Random Forests (RF), which has good noiserobustness and generalization and can well handle unbalanced data classification problem.Moreover, its function that can calculate the importance of variables and its partialdependence function can help to understand the model.This paper first reviews and summarizes predecessors’ research achievements, thenintroduces the theory of RF and its application and research situation. On that basis, this paperstudies the financial early-warning problem of manufacturing listed companies according toproject procedures of data mining, which can be roughly divided into two stages——datapreparation, modeling and evaluation. At the stage of data preparation, the selection principlesof samples and variables are discussed; data distributions are explored by descriptivestatistical analysis and outliers of variables are regulated; abnormal samples in data aredetected and discarded with the help of proximity matrix produced by RF; each variable’simportance is evaluated by RF and the variable combination most contributed to financialearly-warning model is obtained. Compared with the variable combination gained bytraditional variable selection methods, the combination is experimentally verified to be morehelpful for the improvement of the model.At the stage of modeling and evaluation, the optimization of RF’s two main parametersis first introduced; a data experiment is then carried out to testify the noise robustness of RF,whose result shows RF with good noise robustness. With respect to the unbalancedness offinancial early-warning problem, this paper introduces two kinds of methods: one kind isbased on data level——over-sampling and down-sampling, the other is based on algorithmlevel——balanced RF and voting cutoff adjustment RF. Experiment result shows RF based on voting cutoff adjustment is superior to the others. Next, a misclassification cost function isproposed, and RF, SVM and BP neural network are compared by experiment whose resultshows RF has a lowest misclassification cost. Finally, financial early-warning model ofmanufacturing listed companies based on RF is explained by partial dependence function, anda financial distress early-warning region is given to provide references for stakeholders of thecompany.