Loan lending plays an important role in our everyday life. However, loan default is still unavoidable, which carries a great risk and may even end up in a financial crisis. Therefore, it is particularly important for a bank to identify whether a candidate is eligible to receive a loan. In the past, the evaluation primarily depended on manual review, which was time-consuming and labor-intensive. Recently, banks have opted for machine learning approaches to automatically predict loan defaults based on certain features since it can highly enhance the accuracy and efficiency of the prediction.

For my analysis, the dataset is “1999 Czech Financial Dataset - Real Anonymized Transactions” which has been obtained from data.world. It contains real anonymized Czech Bank transactions, account information, and loan records released for PKDD’99 Discovery Challenge.

The relation between the 8 tables is given as:

First datasets are merged based on common columns. Few columns were renamed for the purpose of join() operation. Then columns with more than 50% missing values were removed. Column containing target variable was converted to binary and also converted few categorical columns to numeric format. All the column values were normalized/standardized using scale() as few columns had only single digit values while others had more than 5 digit values. Most important step: In this dataset the number of non – defaulters are 275989 and defaulters are 26262. Under – sampling involves randomly removing instances from the majority class to create a more balanced distribution. This process ensures that the machine learning models are not biased towards predicting the majority class and can better identify patterns associated with the minority class. ovun.sample function from ‘ROSE’ library has been used here.

In this project we created a 80-20 split that is 80% training data and 20% testing data.

For my analysis I considered 3 cases for model development and evaluation for comparison of their performance: Full Model, Reduced model using correlation matrix, and LASSO reduced model.