Predictive modeling in credit risk: some common practices

Abstract

Data mining techniques are used in collecting, cleaning and the initial processing of the data. 80% of the data is randomly selected for building the model and 20% is reserved for validating it. The selected data is then classified into different segments. Segmentation is driven by preliminary analysis and business need. The next step is to study the relationship between the independent and dependent variables for each predictor. Weaker predictors may be discarded at this stage. Transformations of variables are done if needed. After that, stepwise logistic regression is applied on the clean data which eventually produces the model. Model fit statistics are observed. A model that rank orders and displays the best separation from good to bad is considered as the best.