Research On Systemic Risk Of Banking Network Based On Interpretable Machine Learning

Financial security is an important part of national security,among which banking is an important subsystem of financial system and the core of national financial security.The key indicators affecting the stability of the banking network system are studied to explain the specific reasons of the systemic risk of the banking network,so as to reveal the mapping relationship between the parameters of the banking network system and the systemic risk of the banking network,which is the direction of the efforts of countless financial researchers.However,because the actual banking network system has many parameters,strong coupling,and the influence of parameters on the system stability is nonlinear,the traditional analysis method based on computer simulation cannot accurately reveal the mapping relationship between the banking network system parameters and systemic risk.Based on above,the interpretability of banking network system from the perspective of data driven is analyzed in this research,combining with the recently emerging explicable machine learning technology.And the main works of this research are as follows:Firstly,the applications of interpretable algorithm in the systemic risk analysis of banking network are explored,and based on the interpretability of decision tree an analysis method is proposed to analyze the systemic risk of banking network from the macro level.After the decision tree algorithm conducting the secondary modeling on the simulation data of the banking network system,decision tree path and partial dependence plot are used to analyze the systemic risk of banking network.Based on the decision path,the impacts of different parameter combinations on the systemic risk of banking network are obtained.The validity of the interpretable algorithm in the study of systemic risk is verified by comparing with the simulation results when the partial dependence plot is used for single parameter analysis.The premise of interpretability analysis based on quadratic modeling is that the quadratic modeling model has a high fitting performance to the data of banking network system.Therefore,in view of the low accuracy of decision tree algorithm in the secondary modeling of simulation data of banking network system,this thesis proposes an optimization algorithm model based on the algorithm of fitting set tree as a single decision tree to improve the classification accuracy of generating a single decision tree model.Specifically,in view of the unbalanced data set of the banking network system,in the process of training the integrated tree model represented by XGBoost,the traditional cross entropy loss is replaced by Poly Loss loss to improve the classification accuracy of XGBoost algorithm.And in the process of fitting the tree into a single decision tree algorithm,the original simple average method combining the logic of the decision rules is optimized to the weighted average method considering the number of samples of leaf nodes,to further improve the prediction accuracy of the single decision tree algorithm.The results show that the prediction accuracy of the single decision tree model obtained by the improved algorithm is higher than that of the single decision tree.Finally,as the influence of banking network system parameters on systemic risk is coupled and nonlinear,post-interpretation technology is applied to interactive interpretation of banking network system parameters.Specifically,based on Shapley interactive game theory,SHAP computational framework is used to analyze the interaction of two parameters of banking network system on the systemic risk of banking network.The internal mechanism of the influence of the two-parameter interaction on the systemic risk of the banking network is explained in detail,and the reasonable regulation and control suggestions are given.Based on the Shapley value theory,the SHAP framework calculates the contribution value of each feature to the model prediction results,namely the Shapley value.The contribution size of different parameters to the final stability level of the bank system is analyzed through the multi-feature interaction of the decision graph,so as to realize the multi-parameter interaction analysis.