BI APPLICATION IN FINANCIAL SECTOR - CREDIT SCORING OF RETAIL LOANS USING A HYBRID MODELING APPROACH

Retail Loans now-a-days form a major proportion of Loan Portfolio. Broadly they can be classified as (i) Loans for Small and medium Sector and (ii) Loans for Individuals. The objective of Credit Scoring is that we use enough of precaution before the sanction of the loan so that the loans do not go bad after disbursement. This will increase to the bottom line of the financial institution and also reduce the Credit Risk. Techniques used to perform Credit Scoring Varies for the above two classes of loans. In this paper, we concentrate on the application of Credit Scoring for individual or so called personal loans like – Auto loan, buying goods like Televisions, Refrigerators etc. Large numbers of loans are being disbursed in these areas. Though the size of the loan may be small, when compared to Small/Medium Scale Industry, if one does not control the defaults, the consequences will be disastrous. From the characteristics of borrower, product characteristics a Credit Score is computed for each applicant. If the Score exceeds a given threshold loan is sanctioned. If it is below the threshold, loan is sanctioned. If it is below the threshold, loan is rejected. In practice a buffer zone is created near the threshold so that those Credit Scores that fall in buffer zone, detailed investigation will be done before a decision is taken. Two broad classes of Scoring Model exists (i) Subjective Scoring and (ii) Statistical Scoring. Subjective Scoring is based on intuitive judgement. Subjective Scoring works but there is scope for improvement one limitation is prediction of risk is person dependent and focuses on few characteristics and may be mistakenly focusing on wrong characteristics. Statistical Scoring uses hardcore data of borrower characteristics, product characteristics and uses mathematical models to predict the risk. The relation is expressed in the form of an equation which finally gets converted to a score. Subjectivity will be reduced and variable(s) that are important to scoring are identified based on strong mathematical foundation. Different Models have been used in Credit Scoring like Regression, Decision Tree, Discriminate Analysis and Logistic  Regression. Most of the times, a single model is used to compute the Credit Score. This method works well when the underlying decision rule is simple and when the rule becomes complex, the accuracy of the model diminishes very fast. In this Research Paper, a combination of Decision Tree and Logistic Regression is used to determine the weights that are to be assigned to different characteristics of the borrower. Decision Tree is used at first level of analysis to narrow down the importance of Variables and overall weights that needs to be assigned. It is also used for optimum groupings of numeric and non-numeric Variables. At second level, Logistic Regression is used to compute odd ratios a variant of probability, which in turn is used to assign weights for an attribute and to individual levels in an attribute. This has been tested on real life data and found to work better compared to methods using a single stage models. An accuracy of around 80% in decision is obtained which is good for any modeling study as there is no model which gives 100% accuracy. The next Section explains the Methodology, Data Used and Results. SPSS Software has been used for Model Building and Data Analysis