| With the gradual slowdown of the domestic economic growth,the downward pressure on the economy faced by listed companies is also increasing day by day.Manufacturing is a pillar industry in China,and the financial status of manufacturing companies is directly related to national economy.Therefore,it is of great practical significance and application value to establish a set of efficient financial crisis early warning indicator system for listed manufacturing companies to provide early warning to the company’s financial status.At present,the research on financial crisis early warning are very rich,but the following aspects still need to be improved: first,the method of feature selection needs to be improved.In the context of big data,there exists the problem of algorithm non-convergence that the feature dimension is too high and the observation sample size is limited;second,the financial crisis early warning model needs to be improved.Since traditional statistical models often have too strict assumptions and low prediction accuracy;what’s more,the interpretability should be improved.Although the complex ensemble learning model shows high prediction accuracy sometimes,it cannot explain fixed problems.Therefore,this paper starts from these points: firstly,Gibbs sampling is introduced to select features,which can not only describe the nonlinear correlation between features,but also solve the contradiction between insufficient sample size and too many features;secondly,this paper applies the XGBoost model to building a financial crisis early warning model,which can run a large number of tabular data at the same time,and has good accuracy and generalization ability;finally,the SHAP interpretability framework is used to analyze the interpretability of the selected features and deeply study the specific driving effect of each feature on the sample.This paper uses 2116 non-ST listed companies and 125 ST companies in China’s manufacturing industry as the research sample,and determines 76 financial indicators and 14 non-financial indicators to construct an initial financial crisis early warning indicator system.This paper introduces Gibbs sampling to screen the initial index system,and compares it with the traditional feature selection method based on IV value and correlation coefficient.The XGBoost models are constructed according to the respective selected indicators,and the results show that the constructed models using the indicators selected by Gibbs sampling have stronger prediction performance.Finally,with the help of SHAP interpretability framework,this paper analyzes the process by which the indicators affect the prediction results in detail.According to the partial dependence graph of the indicators,it is of great significance to find out the early warning intervals of the indicators for the development of the enterprise.Besides,it is beneficial to provide decision-making basis for the stakeholders. |
