A Thesis Report
CoFounder and MD of PragyanAI
Contact: sateesh.ambesange@pragyanai.com, +91-9741007422
Date: October 2, 2025
Peer-to-peer (P2P) lending platforms have revolutionized the personal finance landscape, but they face the inherent challenge of credit risk management. This project presents a comprehensive data analytics study of the Lending Club dataset to identify the key drivers of loan defaults and develop a robust framework for predictive risk assessment. The study follows a structured methodology encompassing data preparation, in-depth Exploratory Data Analysis (EDA), and the implementation of multiple machine learning models.
Key findings from the EDA reveal that loan grade, debt-to-income (DTI) ratio, interest rate, and loan purpose are the most significant predictors of default. A synthesized profile of a high-risk borrower typically includes a low loan grade (D or worse), a high DTI, and a loan purpose of 'small business' or 'debt consolidation'.
Three distinct classification models—Logistic Regression, Random Forest, and Gradient Boosting—were trained and evaluated. The models demonstrated strong predictive power, with the Gradient Boosting model showing superior performance in identifying potential defaults (high recall). A live simulation tool was integrated into an interactive Streamlit dashboard, allowing for real-time risk assessment of new loan applicants. This report concludes with a step-by-step guide for deploying the application to Streamlit Community Cloud, providing a complete, end-to-end solution for data-driven credit risk management.
In the last decade, P2P lending platforms have emerged as a significant alternative to traditional banking, connecting borrowers directly with investors.
The primary business problem is the financial loss incurred from borrowers who fail to repay their loans (i.e., 'Charge Off'). To ensure platform sustainability and maintain investor confidence, it is critical to identify high-risk applicants before a loan is issued. A proactive, data-driven framework is required.
This report is organized into seven chapters. Chapter 2 details data acquisition and preparation. Chapter 3 presents EDA findings. Chapter 4 describes the predictive modeling phase. Chapter 5 covers the Streamlit app deployment. Chapter 6 provides steps for hosting this flipbook. Finally, Chapter 7 summarizes conclusions.
This analysis utilizes a synthetic dataset mirroring the public Lending Club loan dataset. Attributes include:
loan_amnt, grade, int_rate, annual_inc, dti, purpose, loan_status.A Python script using the pandas library was used for cleaning and transforming data into a usable format.
Sample Code: Basic Data Loading
import pandas as pd
# In a real scenario, you would load a CSV file
# df = pd.read_csv('loan_data.csv')
# For this project, we use a function to generate data
df = create_synthetic_lending_club_data()
print(df.info())
New features were created to enhance model performance:
issue_year: Extracted from issue_d for time-series analysis.credit_history_length: Calculated to represent the borrower's experience with credit.EDA is the process of visualizing and summarizing data to understand its main characteristics and uncover patterns.
The goal was to build a model to classify new applicants as either high-risk ('Charged Off') or low-risk ('Fully Paid').
Three classification models were chosen:
A preprocessing pipeline was built using scikit-learn to scale numeric features and one-hot encode categorical features.
Sample Code: Modeling Pipeline
from sklearn.pipeline import Pipeline
from sklearn.preprocessing import StandardScaler, OneHotEncoder
from sklearn.compose import ColumnTransformer
from sklearn.ensemble import RandomForestClassifier
numeric_features = ['loan_amnt', 'annual_inc', 'int_rate', 'dti']
categorical_features = ['grade', 'purpose']
preprocessor = ColumnTransformer(...)
pipeline = Pipeline(steps=[('preprocessor', preprocessor),
('classifier', RandomForestClassifier())])
pipeline.fit(X_train, y_train)
| Model | Accuracy | Recall (for Default) |
|---|---|---|
| Logistic Regression | 75.3% | 78.1% |
| Random Forest | 88.9% | 65.2% |
| Gradient Boosting | 90.1% | 70.5% |
Gradient Boosting had the highest accuracy, but Logistic Regression was best at catching defaults (highest recall).
Streamlit is an open-source Python library for creating and sharing web apps for machine learning and data science.
The process for deploying the Python dashboard involves a few simple steps:
This interactive flipbook is a static HTML file, meaning it doesn't require a complex server. This makes it a perfect candidate for free hosting services like GitHub Pages.
This study successfully identified the primary indicators of loan default risk, confirming that grade, DTI, and loan purpose are highly predictive.
