LoanTap is a digital lending platform that provides flexible loan products to salaried professionals. With the rise of fintech-driven credit solutions, underwriting accuracy becomes crucial to minimize default risk while ensuring timely loan disbursement. This project builds a predictive model to assess credit risk and assist LoanTap in making data-driven lending decisions.
The objective of this project is to develop a machine learning model that predicts whether a borrower is likely to repay the loan (Fully Paid) or default (Charged Off). The model supports LoanTap’s credit risk team in automating and improving the efficiency of their underwriting process.
The dataset includes borrower-level and loan-level features such as:
- Loan Amount
- Annual Income
- Interest Rate
- EMI
- Credit Score
- Loan Tenure
- Purpose of Loan
- Employment Details
- Repayment Status (Target Variable)
Target Variable:
Fully Paid→ 1Charged Off→ 0
The dataset was imbalanced, with a majority of loans marked as “Fully Paid.”
- Identified skewness and outliers in numeric variables.
- Detected important patterns between features and repayment behavior.
- Handled missing values and ensured clean formatting.
- Encoded categorical variables.
- Normalized numerical features.
- Addressed data imbalance using:
- SMOTE (Synthetic Minority Over-sampling Technique)
- Class Weighting
Built multiple Logistic Regression models:
- Baseline Logistic Regression
- Logistic Regression with Class Weights
- Logistic Regression with SMOTE
- SMOTE + Class Weights
- Threshold-tuned model for best F1-score
- Reglurazed model
- Evaluated using Confusion Matrix, F1-Score, Precision, Recall, ROC-AUC, and PR Curve.
- Tuned the classification threshold using F1 optimization to improve performance on minority class.
- Best Model: Logistic Regression with SMOTE + Class Weighting + Threshold Tuning
- Key Features Impacting Default Risk:
- Zip Code(Geographical presence)
- High EMI relative to income
- Low Credit Score
- High Interest Rate
- Purpose categories like “Debt Consolidation” showed higher risk
- F1-score improved significantly after addressing imbalance and threshold tuning.
- Prioritize 36-Month Loan Terms: Given the higher default rates on 60-month loans, encourage 36-month loans by offering slightly better terms (e.g., lower interest or processing fees) to reduce long-term risk exposure.
- Implement regional risk scoring by incorporating pincode-level default trends. High-risk areas could be subjected to stricter eligibility or additional checks. Limit Loan Size in Risk Bands
- Incorporate external credit bureau data for enhanced accuracy.
- Regularly retrain the model to account for shifts in applicant behavior and economic conditions.
- Experiment with advanced models like XGBoost, Random Forest, and LightGBM.
- Deploy the model using Flask or Streamlit to create an interactive loan approval dashboard.
- Integrate explainability tools like SHAP or LIME for transparent decision-making.
- Monitor model drift and performance using a feedback loop from live loan outcomes.
- You can access the full Python analysis on Google Colab using the following link: View the notebook
A detailed analysis report is available in the following PDF file: View Report.
[SYED ZAHEER ABBAS] - [SYEDZAHEER.C@GMAIL.COM]