International Journal of Computer Applications (0975 – 8887) Volume 161 – No 11, March 2017 1 Credit Scoring using Machine Learning Techniques Sunil Bhatia Computer Science Department VESIT, Chembur Mumbai University Pratik Sharma Computer Science Department VESIT, Chembur Mumbai University Rohit Burman Computer Science Department VESIT, Chembur Mumbai University Santosh Hazari Computer Science Department VESIT, Chembur Mumbai University Rupali Hande Computer Science Department VESIT, Chembur Mumbai University ABSTRACT Lenders such as banks and credit card companies while reviewing a client‟s request for loan use credit scores. Credit scores help measure the creditworthiness of the client using a numerical score. Now it has been found out that the problem can be optimized by using various statistical models. In this study a wide range of statistical methods in machine learning have been applied, though the datasets available to the public is limited due to confidentiality concerns. Problems particular to the context of credit scoring are examined and the statistical methods are reviewed. Keywords Data Mining, Credit Scoring, Logistic Regression, LDA, XGBoost, Random Forest. 1. INTRODUCTION The process of deciding to accept or reject a client‟s credit by banks is commonly executed via judgmental techniques and/or credit scoring models. Earlier, financial institutions and most banks used the method of judgmental approach that is based on the 5 C‟s, which are condition, character, collateral, capital and capacity. In this modern computerized world, this process of deciding can be optimized using statistical methods in machine learning. Thus banks and financial institutions to improve the process of assessing creditworthiness of an applicant during the credit evaluation process develop Credit scoring models. Credit scoring is a system creditors (banks, insurance companies) use to assign credit applicants to either a „„good credit‟‟ group the one that is more likely to repay the debt or a „„bad credit‟‟ group the one who has a high possibility of defaulting on debt or any financial obligation i.e. not paying within the given deadline. Construction of credit scoring models requires data mining techniques. Using, demographic characteristics, historical data on payments and statistical techniques, these models can help in identifying the important demographic characteristics, which is related to credit risk, and assign a score to each customer. The probability that an applicant will default must be calculated from information about the applicant provided at the time of filing the application, and this estimate will thus serve as the basis for his/her creditworthiness. In the paper [1] the four machine learning methods reviewed for Credit scoring jare statistical methods, Hybrid Methods, Artificial Intelligence method, and ensemble learning method. Statistical model includes LDA (Linear Discriminant Analysis), MARS, Decision tree. AI methods include ANN, SVM, and K-Nearest method. Paper also discusses about behavioral scoring method. Behavioral scoring makes a decision about management of credit based on the repayment performance of existing customers during a certain predefined period of time. It also includes repayment behavior and payment history of the client. According to this paper ensemble learning has better prediction accuracy and classification ability and is thus widely applied to personal credit evolution. This paper [2] deals with the design aspects related to financial fraud detection. The aim of feature selection is to improve both the actual and computational performance of the solution, as well as providing a better understanding of the problem. Feature ranking algorithms assign rating to individual features based on certain attributes such as accuracy, content and consistency and choose a suitable subset on the basis of ranking. Performance metrics are used as small increase in performance can lead to large economic benefits. In Classification method accuracy, sensitivity, specificity, precision, false positive rate are the performance measure. In clustering Hopkins statistic is the performance measure. The paper tests various algorithms such as GA1, GA2, DT1, DT2, SVM etc. for determining the best prediction method. It has been found that if misclassification costs are high, techniques with higher sensitivity such as GP1 or neural networks may be suitable choices. If receptiveness to minor chances in dataset is desired then the ant colony optimization or neural networks could be appropriate. Overall the support vector machine could be considered to have the best performance with the highest accuracy. The study applies the credit scoring techniques using data mining of payment history of members from a recreational club [3]. Classification performance of credit scorecard model, logistic regression model and decision tree model were compared. Classification error rates were 27.9%, 28.8% and 28.1% respectively. The cut off score also known as the threshold can be determined by the value of K-S Test for each bucket of score in the validation sample. The target variable is payment status which is a binary variable with 2 categories: default (consisting of persons who have defaulted) and non- default (consisting of person who have not defaulted) which were coded using numerical values (1 and 0). Out of 2765 members, 35% were found to be defaulters. The majority of the members are male (80%) and more than half (74%) of the customers are from non-government sector. Two main limitations are the availability of data and sample selection issues.