Lending Club Loan Data Analysis

A deep learning project for predicting loan default risk using neural networks on Lending Club loan data.

Overview

This project implements a Deep Neural Network (DNN) model to predict whether a borrower will fully pay back their loan or default. The model uses historical loan data from Lending Club, including borrower credit information, loan details, and purpose of the loan.

Dataset

The project uses a loan dataset (loan_data.csv) with the following key features:

• credit.policy: Whether the borrower meets the credit underwriting criteria
• int.rate: The interest rate of the loan
• installment: The monthly installments owed by the borrower
• log.annual.inc: The natural log of the self-reported annual income
• dti: Debt-to-income ratio
• fico: FICO credit score
• days.with.cr.line: Number of days the borrower has had a credit line
• revol.bal: Revolver balance (amount unpaid at the end of credit card billing cycle)
• revol.util: Revolving line utilization rate
• inq.last.6mths: Number of inquiries in the last 6 months
• delinq.2yrs: Number of times the borrower had been 30+ days past due
• pub.rec: Number of derogatory public records
• purpose: Purpose of the loan (categorical feature)
• not.fully.paid: Target variable (1 = borrower will not fully pay the loan, 0 = will fully pay)

Project Structure

.
├── Lending_Club_Loan_Data_Analysis_Guilherme_Faria.ipynb  # Main analysis notebook
├── loan_data.csv                                          # Dataset (not included)
└── README.md                                              # This file

Installation

Prerequisites

• Python 3.7+
• Jupyter Notebook

Required Packages

Install the required dependencies:

pip install pandas numpy scikit-learn tensorflow matplotlib seaborn

Or use a requirements.txt file:

pip install -r requirements.txt

Usage

1. Place the dataset: Ensure loan_data.csv is in the same directory as the notebook.

2. Run the notebook: Open and execute the Jupyter notebook:

jupyter notebook Lending_Club_Loan_Data_Analysis_Guilherme_Faria.ipynb

3. Follow the analysis steps:
   • Data loading and preprocessing
   • Exploratory Data Analysis (EDA)
   • Feature engineering and transformation
   • Model training
   • Evaluation and results

Methodology

1. Data Preprocessing

• Removed missing values and duplicates
• Applied One-Hot Encoding to the categorical purpose feature
• Used StandardScaler to normalize numerical features

2. Exploratory Data Analysis

• Analyzed target variable distribution (16% default rate)
• Visualized class imbalance
• Performed correlation analysis to identify multicollinearity

3. Feature Engineering

• Removed highly correlated features (correlation > 0.80) to reduce multicollinearity
• Final dataset: 18 features after transformation

4. Model Architecture

Built a Deep Neural Network using Keras/TensorFlow with the following structure:

Input Layer (128 neurons, ReLU activation)
    ↓ Dropout (0.3)
Hidden Layer 1 (64 neurons, ReLU activation)
    ↓ Dropout (0.3)
Hidden Layer 2 (32 neurons, ReLU activation)
    ↓ Dropout (0.2)
Output Layer (1 neuron, Sigmoid activation)

Model Parameters:

• Total parameters: 12,801
• Optimizer: Adam (learning rate = 0.001)
• Loss function: Binary Crossentropy
• Batch size: 64
• Epochs: 50
• Validation split: 10%

5. Evaluation Metrics

Due to the imbalanced nature of the dataset, the following metrics were used:

• AUC-ROC Score: Primary metric for imbalanced classification
• Classification Report (precision, recall, F1-score)
• Confusion Matrix

Results

Model Performance

• AUC-ROC Score: 0.6612
• Accuracy: 84%
• True Negatives: 2,389
• False Positives: 25
• False Negatives: 444
• True Positives: 16

Observations

• The model achieves high accuracy but struggles with the minority class (defaulters)
• Low recall for the positive class (3%) indicates the model misses many actual defaults
• The AUC score of 0.66 suggests moderate discriminative ability
• The class imbalance (16% default rate) significantly impacts model performance

Potential Improvements

1. Address Class Imbalance:

   • Use SMOTE (Synthetic Minority Over-sampling Technique)
   • Apply class weights during training
   • Try undersampling the majority class

2. Model Optimization:

   • Hyperparameter tuning (learning rate, batch size, number of layers)
   • Try different architectures (deeper networks, different activation functions)
   • Implement early stopping to prevent overfitting

3. Feature Engineering:

   • Create new features based on domain knowledge
   • Perform more sophisticated feature selection
   • Consider interaction terms between features

4. Alternative Models:

   • Compare with ensemble methods (Random Forest, XGBoost)
   • Try traditional ML models (Logistic Regression, SVM)

Author

Guilherme Faria

Deep Learning Project - SimpliLearn

License

This project is for educational purposes.

Acknowledgments

• Dataset: Lending Club
• Frameworks: TensorFlow/Keras, Scikit-learn, Pandas, NumPy
• Visualization: Matplotlib, Seaborn