Default Prediction Of Listed Companies Based On Sample Weighting And Feature Combination

The default prediction of listed companies is through the construction of the regular relationship between the financial factors,non-financial factors,macro conditions and other feature data of listed companies at t-m time and the state of default at t time,to achieve the use of t time feature data to predict the state of default of listed companies at t+m moment.The following two aspects are studied in this study: First,using feature data of different time windows to predict the default status of new customers may get different results.Generally speaking,the feature data closer to the forecast period has a stronger ability to reveal the predicted default state,and the feature data farther away from the forecast period has a weaker ability to reveal the predicted default state.The second is the selection of the optimal combination of features in the process of predicting the default status of new customers.Using different combinations to predict default status for new customers may also yield different results.This study also has two innovations: first,the sample data are weighted.By giving more weight to the window data closer to the prediction period and to the window data with higher default prediction accuracy,a stochastic forest model is established after weighted averaging the multi-window data.The second is the selection of the optimal feature combination.Feature combination 1 is obtained by the forward selection,feature combination 2 is obtained by backward selection,combination 3 is obtained by taking the intersection of combination 1 and combination 2,and combination 4 is obtained through the union of combination 1 and combination 2.In the feature combination 3 and combination 4,the one with the highest prediction accuracy is selected to get an optimal feature combination.It should be noted that in this paper,the feature combination forecasting accuracy of feature combination 3 and combination 4 is higher than that of feature combination 1 and combination 2.It is found that there are 9 indicators with the ability to predict default of Chinese listed companies for 0-5 years,such as "asset-liability ratio" and "current ratio",and 30 indicators with short-term forecasting ability for 0-2 years,such as "asset-liability ratio","growth rate of total exports(US $100 million)","EVA ratio of net assets" and so on.There are 24 indicators with medium-and long-term forecasting ability of 3-5 years,such as "asset-liability ratio","earnings reserve per share","unemployment: growth rate of unemployment rate(%)" and so on.In the one-year forecast period,16 indicators such as "comprehensive leverage","asset-liability ratio","industry climate index","current ratio" and "retained earnings per share" have a key impact on the default prediction of listed companies.Among the top 10 indicators,there are five financial factors,namely,"asset-liability ratio","current ratio","ratio of capital reserve to owners’ equity","comprehensive leverage" and "retained earnings per share".There are 4 external macro condition indicators,they are "Industry Prosperity Index","resident consumption level Index: all residents(last year = 100)","growth rate of Railway operating mileage(10,000 km)" and "growth rate of total afforestation area ".The research shows that the prediction accuracy of "AUC","G-mean" and "Recall" of the default prediction model of listed companies based on random forest is better than that of typical deep learning models such as artificial neural network(ANN,Artificial Neural Network),deep neural network(DNN,Deep Neural Network),machine learning model such as decision tree model(DT,Decision Tree),K nearest neighbor model(KNN,K-Nearest Neighbor),statistical model such as logical regression model(LR,Logistic Regression).