Neural Network Credit Scoring:Evidence From Business Applications

With the marvelous growth of the credit industry and the diversified loan portfolios,credit scoring has gained more and more attention as the credit industry can then benefit from reducing the possible credit risks,improving cash flow,ensuring credit collections and enhancing the better managerial decisions.The accuracy of credit scoring is critical to financial institutions' profitability.Even a tiny percentage of improvement on the credit approval modeling accuracy of credit decision will produce a momentous potential reserve for financial institutions.Therefore,the ultimate goal of credit scoring models is to assign credit applicants to either a 'good credit' group that is likely to repay financial obligation or a 'failed credit' group whose claim will be refused because of its high likelihood of defaulting on the financial obligation.Under such an environment,the objective of this dissertation is particularly on credit scoring and modeling credit approval data,to advance the model performance and perk up the forecasting precision.When we look at the last two decades,Artificial Neural Networks(ANNs),a fashionable credit scoring techniques comes out as an important alternative and draws attention from many researchers with its high forecasting accurateness.Although neural network are gradually commenced to be more influential in many forecasting and classification application,the recital is essentially related to the network algorithm itself,especially on initial condition,network topologies and training algorithms,which may be one reason why the results of neural network for credit scoring appraisal varies when compared with different architectures,with different traditional models and even with ensemble classifiers.To get the most excellent neural algorithm is still a challenging agenda.Furthermore,the potentiality of the classifiers depends on the details of the problem,the data structure,the characteristics used,the extent to which it is possible to segregate the classes by using those characteristics,and the objective of the classification.No single forecasting classifier can generate the optimum accurateness for all classification and forecasting problems.As a result,there is an increasing propensity that traditional applications of an individual algorithm can make an additional development by the hybrid method through feature engineering.Therefore,the core objective of the first empirical study is to inspect the prediction performance of hybrid classifiers by comparing single,advanced statistical and individual neural classifiers.Particularly,the combination of the feature engineering with popular neural network classifiers;a hybridization approach,is compared with a hybrid classifier,single neural classifiers,and three well-known baseline classifiers.Overall,we executed a 12 + 8 +(8 × 8)experimental design that resulted in 84 unique classification models,being examined over a large credit approval example set from a Chinese commercial bank.However,the design of consistent classifiers to forecast credit granting choices is critical for many financial decision-making practices.Although there are some artificial and statistical techniques,have been developed to predict customer insolvency,how to provide an inclusive appraisal of prediction models and recommend adequate classifiers is still an imperative and understudied area in credit approval data modeling problems.Previous evidence points out that ranking of classifiers varies on different criterions with their measures under different circumstances.Therefore,in second empirical study,this dissertation addresses this methodological flaw by proposing simultaneous application of probabilistic neural network(PNN)and support vector machine(SVM)based credit scoring algorithms,together with a frequently used high-performance models,and accordingly fill this gap by introducing a set of multi-dimensional evaluation measure being combined with some novel metrics which would be helpful in discovering unseen features of the model's performance.For the effectiveness and feasibility purposes,six real-world credit datasets have been applied.Empirical study shows that PNN model is more robust than its rivalries and traditional performance evaluations are about consistent with their original counterparts.With these contributions,therefore,our investigations bring several advantages to financial risk management.Then,the heart of the third empirical study is particularly on risk assessment of financial decision support systems(FDSSs),to advance the model performance and improve classification accuracy.To conquer the downsides of the classical models,Artificial Intelligence(AI)technologies,e.g.,multilayer perceptrons(MLPs)and support vector machines(SVMs)have been deliberated for FDSS applications.Recently the prestigiousness of AI approaches has been confronted by the latest prediction learners.Therefore,to ensure the competitive performance of AI mechanisms,the current investigations scrutinize the topological applications of MLPs and SVMs over eight different databases with equivalent combinations in credit approval and bankruptcy predictions example sets.The experimental results reveal that MLP5-5 and MLP4-4,i.e.,the sigmoid activation function with 5 and 4 hidden layers are the feasible topologies for MLP algorithm;then on all databases in all performance criterions SVM trained with the linear kernel function(SVM-1)achieved better prediction results;from the 'Baseline' family,random forest(RF)learner has brought significant improvements in financial decisions;and FDSSs are correlated with the nature of databases and the performance criterions of the trained algorithms.With these contributions,therefore,we supplement earlier evidence and enhance the predictive performance of AI algorithms for managerial decision support applications.The results of this dissertation,however,have practical and managerial implications on FDSSs.It will guide to the execution of risk-adjusted loan pricing systems,being supplemented some add up to the financial literacy of the investigated study.Corporate practitioners could able to set up the feasible regulatory capital requirements from the findings and methodological schemes of this project.This study will also aid the credit lenders and borrowers to protect themselves from the potential clients with high credit risk in a timely manner,which would facilitate to make a range of financial and non-financial decision-making strategies.Given that financial and operational resources are limited,findings gained from this study and its methodology could assist for feasible segmentation of good credit customers and target the scarce resources for better arrears management.