Farmer Microfinance Credit Rating Model Research Based On The Default Discrimination

The credit rating refers to whether the debtor can repay the principal and interest on time.Its essence is the "discriminating ability of the default state",which is used to identify the debtor's default risk.The accuracy of the credit risk assessment method is critical to the overall social economy,and the 2008 financial crisis was caused by credit risk management errors.Conversely,a good credit risk identification system will help banks reduce large losses.China is a big agricultural country,but the imperfect financial market in rural areas seriously restricts the development of rural economy.Farmers generally have large capital needs for production,but farmers lack the characteristics of collateral,which leads to difficulties in financing.Therefore,we need to build a reasonable credit rating system for farmers' microfinance,which is the key to solving the financing difficulties of farmers.The thesis takes a sample of farmer's microfinance data from 28 provinces in a national commercial bank.We take "the most discriminating ability of default status" as the core,and carry out research on the logical main line of "selection of credit rating indicators ?construction of incremental credit scoring model division of credit ratings".The credit rating study on farmer microfinance is divided into the following three parts.First,we carry out the feature selection of credit rating,consider the overall default discriminating power after the mutual influence of indicators,rather than the default discriminating power of individual indicators,and construct an indicator system based on the maximum discriminative power of default status.Second,we built an incremental credit scoring equation.After constructing the initial credit scoring equation based on the old data,when the new data is reached,we can no longer use the old data,and only use the key parameters in the initial credit scoring equation and the new data,update the credit score equation.Third,we divide the customer into different credit ratings,and classify the credit according to the two objective functions of loss rate monotonous distribution and customer number distribution,so that the credit rating result satisfies the monotonicity of the loss rate.The above three parts correspond to Chapters 2-4 of this thesis,and these three parts are gradually progressive.Chapter 2 Feature Selection provides the basis for the features in Chapters 3 and 4;Chapter 3 Credit Scoring is based on the features selected in Chapter 2,and the credit score data is provided for Chapter 4.Chapter 4 Credit Rating is to divide farmers into different credit ratings based on the credit score of Chapter 3.The main works and innovations of this thesis are as follows:(1)The innovation of feature selection:We improve the existing sequential backward search algorithm the by add the two steps of "backtracking the deleted indicators" and"replacing the weak feature" in the process of sequential backward search algorithm.The method proposed in this thesis improves the existing sequential backward search method which ignored many feature combination and resulted into the weaker discriminant power.Finally,we constructs a credit rating index system with strong default discriminant ability.The thesis proposes an improved backward floating selection algorithm(IBFS),which adds two steps based on the sequential backward selection algorithm.First step is backtracking,which attempts to use sequential selection method to add one feature from the deleted features.Checking if backtracking helps improving AUC,then add the feature.Second step is replacing,which is to check whether removing any feature in the currently selected feature subset and adding a new one at each sequential step can improve the current feature subset,if replacing helps improving AUC,then replace the weak feature.Through the process of backtracking and replacing,the missing feature subsets are reduced to improve the criteria AUC.The proposed method improves the drawback that the existing sequential backward search method neglects many feature subsets,and ensures that the selected features can significantly identify the default status of the farmer's microfinance.(2)The innovation of credit scoring:When the new data is reached,the old data is no longer used,only the key parameters in the original credit scoring equation and the new data are used to update the credit scoring equation.We have changed the shortcoming that existing research don't re-learning when new data arrives,or re-do batch learning every time based on the all data combined the old and new data.Finally,we constructed a credit scoring equation based on the incremental extreme learning machine that can be dynamically updated.First,the SMOTE algorithm is used to synthesize new default samples,to solve the imbalance problem that samples consist of excessive sample of defaulted farmers and excessive sample of non-defaulting farmers.Secondly,for the old sample data,the connection weight vector between the hidden layer and the real default state is solved by introducing the generalized inverse matrix,so as to construct the initial credit scoring equation ?k=1L?kN0G(wk,bk,xj),which avoids the cumbersome process of the traditional neural network through the complicated iterative solution.Finally,when the new data Dnew={(xj,yj)}arrives,the key parameters ?kN0 in the extreme learning machine credit scoring model obtained from the old data and the new data are used to update the parameter ?kN0+Nnew in credit scoring function,in order to determinate the new credit scoring function?k=1L?kN0+Nnew(wk,bk,xj).To ensure that the credit scoring model can be continuously updated while the data increases,so that the credit scoring model can maintain a high classification accuracy of default farmers,avoiding the need to completely re-learn based on"new + old" all data while each new data arrives,and avoiding to use old credit scoring model without considering the impact of new data on the credit scoring model.(3)The innovation of credit rating:we constructs a non-linear credit rating model by controlling the monotonicity of loss rate and bell-shaped distribution of the number of customers,which improves the drawbacks of ignoring the distribution of customer numbers and the monotonicity of the loss rate.According to the pyramid criterion that means the higher the credit rating,the lower the loss given default,and considering the fact in finance that the very few clients show either excellent or very low credit quality,most obligors have an average credit quality.We propose a credit rating model that establish non-linear multi-objective programming by minimizing the accumulated difference between the proportion of customers in each grade and their expected proportion under a normal distribution,and minimizing the difference in loss rate between any two adjacent grades.Then we apply genetic algorithm to solve the model.The proposed credit rating model will ensure that the number of clients for each rating follows a bell-shaped distribution,and the loss rate is monotonic.The proposed credit rating model can provide the key parameter in loan pricing to cover the credit risk,and improves the drawback that existing researches ignore the customer distribution which often induce a change of credit grade following small changes in the credit score.The main conclusions of this thesis are as follows:(1)The conclusions on the feature selection:This paper empirically demonstrates the data of 2044 rural microfinance loans of a national commercial bank.In the end,this paper constructs a credit rating index system for farmers' microfinance loans including 17 indicators such as "marital status","supporting population" and "skills status of lenders and their families".These features cover the basic situation,repayment ability,repayment willingness,guarantee of joint guarantee,and macro environment.The default discriminative power AUC of the farmer's microfinance credit rating index system constructed in this paper is 0.7291,which is better than other methods such as sequential backward search algorithm.(2)The conclusions on the incremental credit scoring model:First,when the ratio of"number of defaulted samples:number of non-defaulting samples" is "1:2",the highest classification accuracy is determined.Second,the thesis uses the KS curve to verify the credit score equation,the critical point of the best credit score is determined to be 40.74 points,which can maximize the distinction between default and non-defaulting farmers.Third,the credit scoring model based on the incremental extreme learning machine method has a higher ability to discriminate default status,especially in the discriminant accuracy of default farmers,which is superior to the stochastic gradient descent(SGD)model and other methods.(3)The conclusions on credit rating:The results of the proposed credit rating model not only satisfy that the number of clients for each rating follows a bell-shaped distribution,and the loss rate is monotonic.This avoids the irrationality of most samples gathering around AAA or C rating.The credit score interval obtained in this study is relatively stable,which makes up for the insufficiency of the credit grade change caused by the small change of the credit score of the existing research.