Credit Scoring Model's Performance Evaluation And Credit Risk Evaluation Of Small Enterprises

Micro,small,and medium enterprises(MSMEs)are an important part of the Chinese economy,and have become important drivers of social progress.The Chinese government has implemented various policies to support the growth of private companies,including credit increases and reductions in taxes and fees.Nevertheless,access to financing remains a challenge for MSMEs.There are several characteristics of MSMEs that may result in the unwillingness of banks to build a credit relationship with MSMEs,including the information opacity problem,the high cost in acquiring the ”soft information”of MSMEs,and the fact that MSMEs are more vulnerable to changes in the external environment.The credit risk of MSMEs is higher than large enterprises and may change over time due to these factors,which makes it difficult for financial institutions such as banks to access the credit risk of these firms accurately and timely,based on static historical information such as financial statements.From banks' perspective,maximizing the profit and minimizing the risk are two objectives they may consider when making a financial decision such as granting loans.When the majority of borrowers are MSMEs,the cost parameters that quantifying model's economic losses caused by misclassifications of defaulters and non-defaulters are often uncertain due to the dynamic characteristic of the MSMEs' credit risk,which makes it impossible to estimating the model's profit precisely by using previous measures.Besides,the cost parameters uncertainty results in the uncertainty of future returns for financial institutions using the credit scoring model,and may lead to severe losses during worst-cases.Although numerous profit-based measures,such as the maximum profit measure,have been proposed in credit scoring,quantifying credit scoring models' economic profit and risk with uncertain cost parameters remains a challenge.With the development of data-storage technology,banks have accumulated a large amount of transactional data.Those transactional data are timely updated,reliable and reflect MSMEs' economic activities.This dissertation proposes a credit risk model for MSMEs based on the transactional data provided by banks,to access the asymmetric information problem and evaluate MSMEs' credit risk timely.Further,this dissertation proposes a robust profit measure to evaluate the model profit,and introduces the robust risk measure to estimate the model risk under the cost parameters' uncertainty.Based on the new metrics,we propose a feature-selection framework that optimizes model profit and model risk simultaneously to support model development.We list the main contents of this thesis as follows.First,we propose two metrics to access the model's profit and risk under cost parameters' uncertainty respectively.Due to the dynamic of MSMEs' credit risk,estimating the exact value(or the exact distribution function)of cost parameters for historical data is difficult,if not impossible.We propose a worst-case expected minimum cost measure(WEMC)under the parameters' uncertainty.WEMC provides a conservative estimation of model profit under the ambiguity of cost parameters.Meanwhile,we exploit the worst-case conditional value-at-risk(WCVa R)measure for estimating the model loss when extreme cases occur.Furthermore,we investigate the correlations between new metrics and traditional model performance measures through theoretical argument and empirical analysis.Second,we construct a credit risk model for MSMEs using transactional data-based variables to access the information opacity and the high default risk characteristics of MSMEs.Compared with financial statements and soft information gathered by contacting over time with the firm,transactional data are constantly updated,quasi-costless,reliable and of high quality.Using data provided by Shandong City Commercial Bank Alliance Co.,Ltd.,we construct transactional data-based variables to estimate the credit risk of MSMEs.Meanwhile,MSMEs do not live in isolation,but connected with each other by daily transactions and business partnerships.Those daily interactions are essential for MSMEs' business operation.Therefore,we also take the payment interactions among MSMEs into consideration when evaluating the credit risk of MSMEs by incorporating payment network-based variables into credit risk prediction.Empirical results demonstrate the predictive power of transactional data-based and payment network-based features.Third,we propose a profit and risk-driven feature-selection framework to increase the interpretability of the credit scoring model.Incorporating MSME payment information in credit risk evaluation improves the model performance,but increases the cardinality of feature set,which is adverse to model interpretability.Besides,the business objectives—maximizing the profit and minimizing the risk—are rarely considered during model deployment.In this dissertation,we propose a two-stage multi-objective feature-selection framework.In the first stage,we eliminate irrelevant and redundant features by filter method.In the second stage,wrapper-based multi-objective feature-selection methods that optimize model profit and model risk simultaneously are applied to find the optimal feature subset.Empirical results show that the proposed methods greatly reduce the dimension of data and therefore increase the interpretability of the model.Furthermore,the proposed methods provide a return-risk trade-off for decision-makers as the solution allows decision-makers to select a proper feature subset based on their risk preference.Finally,we analyse the importance of features in the selected feature subset,the results confirm the importance of transactional data-based and payment network-based features in improving model profit and decreasing model risk.Based on the characteristics of MSMEs' credit risk,we propose two metrics,i.e.,profit-based and risk-based model performance measures,to measure the performance of a credit scoring model,and construct a credit scoring model using transactional data.The proposed metrics add to profit-based classification model performance measures by taking the uncertainty of cost parameters into consideration.Furthermore,we estimate the model risk under the cost parameters' uncertainty and provide a quantifiable criterion for model risk management,which to the best of our knowledge,no previous study has attempted to do so.Besides,we demonstrate the empirical value of the proposed model on the offline and online test environment provided by Shandong City Commercial Bank Alliance Co.,Ltd..Finally,we propose a two stage multiobjective feature-selection method to access credit scoring model construction.The proposed feature-selection framework that simultaneously optimizes model interpretability,model risk and cost in model development,which is aligned with banks' objectives.Meanwhile,the proposed methods output a Parato non-dominated solution set and therefore allow banks to select the optimal feature subset according to their risk preference.