Loan Default Prediction Model

Overview

This project develops a machine learning model to predict loan defaults using historical loan data from Lending Club. The primary goal is to build a scorecard that assigns risk scores to loan applicants, aiding in the assessment of their creditworthiness. The project encompasses data preprocessing, feature engineering, model selection, and out-of-time validation to ensure robust and generalizable performance.

Data

The dataset used in this project is obtained from Lending Club. It includes information about accepted loan applications, such as loan amount, interest rate, borrower demographics, and loan status.

• accepted_2007_to_2018Q4.csv: Contains accepted loan applications data from 2007 to Q4 2018.
  • A subset of this data, specifically loans issued in August-December 2014, is used for training.
  • Loans issued in Jan-Mar 2015 are used for Out-of-Time validation
  • Dataset : Kaggle
  • Unzip and save in a folder ./data

Key Steps

1. Data Loading and Preparation:

   • Loads loan data from CSV files.
   • Creates a binary target variable (Bad_Flag) indicating loan default.
   • Subsets the data for training and out-of-time validation.

2. Exploratory Data Analysis (EDA) and Feature Engineering:

   • Analyzes categorical and date variables.
   • Drops irrelevant columns.
   • Creates the target variable 'Bad_Flag' based on 'loan_status'.

3. Missing Value Imputation:

   • Identifies and handles missing values using a tiered approach:
     • Drops variables with >= 80% missing values.
     • Imputes variables with 50-80% missing values with 99999 (or 0 in OOT validation due to data differences).
     • Imputes variables with < 50% missing values using a bad-rate informed median imputation.

4. Outlier Handling:

   • Caps outliers in numerical variables at the 5th and 95th percentiles.

5. One-Hot Encoding:

   • Converts categorical features into numerical features using one-hot encoding.

6. Feature Selection:

   • Calculates KS statistic and Gini coefficient for each numerical variable.
   • Performs Pearson correlation analysis.
   • Manually selects a subset of variables based on statistical significance and business intuition.
   • Calculates Variance Inflation Factor (VIF) to check for multicollinearity.

7. Model Training and Evaluation:

   • Explores three classification algorithms: Ridge Regression, Random Forest, and Gradient Boosting Machine (GBM).
   • Splits data into training and testing sets.
   • Trains each model on the training data.
   • Evaluates models using accuracy and log loss.

8. Out-of-Time (OOT) Validation:

   • Tests the model on a separate OOT dataset (January-March 2015 loan applications).
   • Calculates accuracy and log loss on the OOT data.

9. Scorecard Development:

   • Develops a scorecard based on the predicted probabilities from the Random Forest model.
   • Scales probabilities to create a score range from 300 to 850.

Code Structure

The project code is organized as follows:

• Import Libraries: Imports necessary libraries, including NumPy, Pandas, scikit-learn, statsmodels, matplotlib, and seaborn.
• Data Loading and Preparation: Reads and preprocesses the loan data.
• EDA and Feature Engineering: Explores the data and engineers relevant features.
• Missing Value Imputation: Handles missing values using the described methods.
• Outlier Handling: Caps outliers to reduce their impact.
• One-Hot Encoding: Converts categorical variables to numerical format.
• Feature Selection: Selects relevant features using statistical and business criteria.
• Model Training and Evaluation: Trains and evaluates Ridge Regression, Random Forest, and GBM models.
• OOT Validation: Validates the model on out-of-time data.
• Scorecard Development: Creates a risk scorecard based on model predictions.

Libraries Used

• numpy: Numerical computing
• pandas: Data manipulation and analysis
• matplotlib: Data visualization
• seaborn: Statistical data visualization
• scikit-learn: Machine learning algorithms and tools
• statsmodels: Statistical modeling and econometrics
• scipy: Scientific computing

Results

• Model Performance: Random Forest achieved the best performance with a test accuracy of 98.41% and an OOT accuracy of 98.36%.
• Feature Importance: Key predictors of loan default include recoveries, last_fico_range_high, total_rec_prncp, last_pymnt_amnt, and int_rate.
• Scorecard: The developed scorecard assigns risk scores from 300 to 850, providing a practical tool for assessing loan applicant risk.

Algo	Train_Error	Train_Acc	OOT_Error	OOT_Acc
Ridge	0.0822002245989056	0.9662	0.08556710142894697	0.9641
GBM	0.07316159106628008	0.9753	0.07599692128878723	0.9741
Random Forest	0.05165800250375114	0.9841	0.06180492322965778	0.9836

Usage

1. Clone the repository:

   git clone https://github.com/Prad06/loan-default-prediction
   cd loan-default-prediction
   

2. Install the required libraries:

   pip install numpy pandas matplotlib seaborn scikit-learn statsmodels scipy
   

3. Run the code:

   jupyter notebook loan-default-prediction.ipynb
   

Next Steps

• Further refine feature selection to improve model interpretability and reduce complexity.
• Explore additional machine learning algorithms and hyperparameter tuning techniques.
• Implement the scorecard in a production environment for real-time loan risk assessment.
• Incorporate external data sources to enhance the predictive power of the model.