Credit Default Prediction Model Based On Support Vector Machine And Its Applications

Credit approval data modeling is a significant research agenda in the credit industry.With the fast expansion in this field,credit default prediction(CDP)classifiers have been extensively used for the credit customers' assessment and execution of large loan portfolios.Particularly,default risk prediction data modeling is the binary classification problem in the context of pattern recognition theory,which aspires at conveying new observations to predefined decision classes.Two primary concerns have thoughtful impacts on the outcome of the credit risk decision classes:selecting an accurate feature selection algorithm to find the reasonable feature set and choosing the right classification model to construct the decision class using that feature set.Managing financial risk is one of the most insightful issues that should inspect to explore many significant parameters.The banking industry comprises numerous risk factors influencing both banks and their stakeholders.CDP has an intense association with banks being an efficient and decisive technique for experimenting with the borrowing and lending of money.It is important to accmulate data about creditors,regulators,other financial and non-financial companies,government to supervise the credit risks.Simultaneously,CDP is also important to arrive at a central assessment to lend money to their customers.Moreover,the CDP approach can assist in segregating creditworthy customers from their non-creditworthy counterparts.It means that some credit customers have spotless and good evidence;accordingly,banks can categorize them as "solvent creditors." On the contrary,a few others,not carrying such clean grounds,and thereby classified as "insolvent creditors."However,it is worth noting that such a straightforward classification procedure may not provide an optimum credit risk management process.Therefore,novel precise automated systems enhancing the prediction accuracy immediately require managing large and multifaceted credit approval instances.Based on this background,the objective of this dissertation is to appraise the credit customer's default risk prediction with an automated support vector machine(SVM).Linking to this aim,the first empirical study(chapter three)aspires to focus on feature selection methods for support vector machines(SVMs)classifiers and check their optimality by comparing results with some statistical and baseline methods.It exploits seven customary individual feature selection methods from the family of filters and embedded approaches by splitting a Chinese credit approval database.Then as an extension of feature selection methodology,the second empirical study(chapter four)includes five 'new age' group feature selectors that are not applied in previous literature in predicting loan customers' creditworthiness.These are the least absolute shrinkage and selection operator(Lasso),ridge regression(RR),random forest(RF),gradient boosting(GB),and least angle regression(LAR).Moreover,chapter four also assembles composite with representative performance metrics;those are combined with traditional and new measures that evaluate the quality of default prediction classifiers.For the testing and illustration purposes,chapter four applies six real-world credit approval diversified databases over six different classifiers.After that,the third empirical study(chapter five)enhances the default risk prediction methodology by applying hybrid approaches.It proposes LogitSVM(logit regression+SVM;LSVM)and LogitNeural(logit regression+neural network;LNA)as a new blending credit risk assessment algorithms.Nevertheless,chapter five applies three real-world credit databases to validate the usefulness and feasibility of the proposed risk assessment blending approaches.Therefore,the first empirical findings suggest that conventional t-test(SVM1)in training-testing sets and DTQUEST(SVM7)in 'no sample division' show outstanding recital in respect of all performance criteria.Moreover,the average results from sample division achieved a more robust prediction ability than 'no sample division' instances.Then,the second empirical findings approve that the ridge regression-based 'new age'feature selection algorithm is superior to Lasso and GB,being more robust than its other counterparts.Then the SVM approves a more suitable classifier as an individual learner.Eventually,the third empirical findings reveal that on the first hand,the blending LogitSVM is efficient than its peers LogitNeural classifiers;on the other hand,kernel-based SVMs also outperform both LNA and back-propagation neural network(BPN).From the contribution point of view,firstly,this study establishes an optimal individual feature selection methodology.It significantly discriminates the non-default credit customers from their default counterparts.It provides a comprehensive study of different individual feature selection methods.Moreover,the current dissertation differs from existing studies for which no comprehensive feature selectors having been applied like this dissertation.In addition,the support vector machine solves the multivariate data normality problem that exists in statistical models.This study,however,has practical significances for financial institutions,managers,employees,investors,and government officials to sort out forthcoming lending transactions to attain targeted risk-return tradeoff,which helps to minimize risk in their decisions.The second contribution is that this study ascertains an optimal feature set selection methodology.It differs from the first contribution as the earlier one deals with one by one feature,which is a traditional methodology.The group feature sets significantly enhance the default risk prediction data interpretation performance by increasing accuracy,AUC;and decreasing type ? error and type ? error as well.Moreover,this study differs from the existing studies,as it selects the best 'new age' feature selectors along with the new performance measures and baseline models.As well,the 'new age' feature sets solve the drawbacks of customary feature selection techniques and enhance the overall default risk discrimination ability.Eventually,the study contributions might be helpful for allocating capital,identifying the priorities for scrutiny,and upgrading the institutional performances.The final contribution is that this study establishes a feasible hybrid default risk prediction technique.The mixtures of LogitSVM and LogitNeural boost up the credit risk discrimination ability by ensuring the model's augmentation,diversity of prediction assignments,and multi-functionality.The hybrid default risk prediction approaches differ with a single classifier because the former one ensures more profitability and cost-effectiveness.Alongside,the hybrid techniques applied in this dissertation solve the overfitting issues of other studies.Accordingly,it improves default risk discrimination ability.Moreover,this dissertation establishes the most scalable fusion techniques which may assist in reducing the enormous amount of monetary losses for the financial industry.Regarding the significance,the experimental settings and numerical results of this study can not only assist the stakeholders and the audience of this dissertation to know the depth scenarios of hybrid credit risk assessment issues,but also it has varieties of practical and managerial schemes in different application areas.Concerning the policy implication and managerial applications,this study supplements earlier literature and update the methods of credit risk assessment modeling.Moreover,this dissertation brings several advantages to credit risk management,including default prediction.Firstly,the proposed CDP modeling methodology can use to recognize the best useful business practices,innovative ideas,and effectual operating measures with successful winning strategies that can be approved by financial institutions to speed up their progress by ensuring productivity,quality,and cost improvements.Secondly,the prediction results can also apply to carry out other objectives,for example,allocating finance or identifying the priorities for scrutiny and upgrading the institutional performance;and these all advantages will enhance the credit repayment,decrease possible credit losses,an automatic increase of profits which help to formulate successful managerial decisions.Finally,it hopes that policymakers have the opportunity to investigate customer financial practices of the peer groups and that they are able to increase their future competence by adapting these methodological issues for their probable default customers.