Predicting Default Risk - Creditworthiness

In this case, we are going to predict the default possibility of the new 500 loan applications to process this week. This is a project from Udacity - Predictive Analytics for Business Program.

Step One: Prepare Dataset

After using a Field Summary Tool, we get a glance of our training dataset. We removed some of the variables due to a large of number of missing data. For the variables with little missing numbers, we filled up the Nulls with the median of the filed.

Based on Pearson Correlation Analysis result, we filtered three predictor variables significantly correlated with target variable.

In order to clean original data, please take a look at the workflow below.

Step Two: Train Classification Models

Create Estimation and Validation samples where 70% of the dataset goes to Estimation and 30% of the entire dataset is reserved for Validation. Here, we trained the following models: Logistic Regression, Decision Tree, Forest Model, Boosted Model.

In terms of overall model accuracy, Forest Model is the highest, while Decision Tree Model is the lowest. With respect to Creditworthy accuracy, both Forest model and Boosted model have the same highest Creditworthy accuracy 0.9619, while Decision Tree model has the lowest 0.8667, and Stepwise LR model has a slightly higher accuracy 0.8762 compared to Decision Tree model.

In respect of Non-Creditworthy accuracy, Stepwise Logistic Regression model has the highest Non-Creditworthy accuracy 0.4889, while Boosted model has the lowest Non-Creditworthy accuracy 0.3778.

From ROC curve, Forest Model’s true positive rate is slightly higher than other.

Step Three: Score Forest Model

Conclusion: After we score our “customer-to-score” dataset, we add a Filter Tool to count the number of individuals, where Score_yes is greater than Score_No. We find 409 individuals are creditworthy within the 500 new applications.