Loan Credit Worthiness Prediction Model

Project Overview

This project aims to predict the creditworthiness of individuals applying for loans. By analyzing financial and personal data, the goal was to develop a model that could predict whether an individual is likely to repay or default on a loan. This project offered some unique challenges, particularly with skewed and imbalanced data, but I learned a lot throughout the process.

Dataset Overview

The dataset consists of information about applicants, their financial history, and personal details. Below are the columns present in the dataset:

• ID: A unique identifier for each observation.
• Loan_ID: A unique identifier for each loan.
• Gender: Applicant’s gender (encoded as 0 for male and 1 for female).
• Married: Marital status (0 = Not Married, 1 = Married).
• Dependents: Number of dependents relying on the applicant, ranging from 0, 1, 2, to 3+.
• Education: Education level of the applicant (0 = Undergraduate, 1 = Graduate).
• Self_Employed: Employment status of the applicant (0 = No, 1 = Yes).
• ApplicantIncome: Monthly income of the applicant, ranging from 0 to 80,000.
• CoapplicantIncome: Monthly income of the co-applicant, from 0 to 20,000.
• LoanAmount: Loan amount requested in thousands (range: 20 to 800).
• Loan_Amount_Term: Loan repayment term in months (range: 0 to 500).
• Credit_History: Indicates whether the applicant’s credit history meets certain guidelines (0 = No, 1 = Yes).
• Property_Area: Type of area where the applicant lives (encoded as 0 = Rural, 1 = Semi-Urban, 2 = Urban).
• Loan_Status: Whether the loan was approved (Y/N); this is the target variable and only present in the training set.
• Total_Income: Total income of the applicant and co-applicant, ranging from 0 to 20,000.

Key Insights

One of the first things I noticed was the skewness and class imbalance in the dataset. The majority of applicants had their loans approved (Class 1), while a smaller portion had their loans denied (Class 0). This imbalance posed a challenge for model performance, as the model tended to predict the majority class (approved loans) more often. To address this, I opted for class weighting instead of using SMOTE, which caused some issues in my initial trials.

Data Preprocessing

In order to prepare the data for modeling, several key steps were taken:

• Handling Missing Values: Missing data was imputed based on available information.
• Feature Engineering: I added Total_Income, a combined metric of the applicant and co-applicant's income.
• Class Weighting: Due to the imbalanced nature of the target variable, I applied class weighting to give more importance to the minority class (loan denials).
• Normalization/Scaling: I normalized certain features like ApplicantIncome, LoanAmount, and Total_Income to make them comparable.

Model Selection & Evaluation

The Elastic Net model performed the best among the various models I tried. Here are the key performance metrics:

• Recall for Class 0 (Loan Denied): 0
• Precision for Class 0: 0
• Recall for Class 1 (Loan Approved): 1
• Precision for Class 1: 0.83
• Accuracy: 83%

Although the recall for Class 0 (loan denial) was 0, the model was able to accurately predict the majority class (loan approval), which is critical for risk management in loan approval processes.

Challenges

• Class Imbalance: The dataset was heavily imbalanced, with more approved loans than denied loans. Handling this imbalance was a key challenge.
• Feature Selection: Some features like CoapplicantIncome were less predictive, and understanding which features to keep or discard was essential.
• Model Performance: Despite experimenting with multiple models, Elastic Net was the most promising due to its ability to handle the imbalanced dataset.

How to Use

To run the model, follow these steps:

1. Clone the repository:

   git clone <repository_url>