After exploring millions of loan records from Lending Club, we found that people with higher interest rates, lower credit scores (FICO), and worse loan grades are way more likely to default. Among the models we tested, XGBoost worked best at predicting this, with a ROC AUC of 0.72 — not perfect, but pretty solid.
DataCents is a collaborative research team using data science to decode credit risk in peer-to-peer (P2P) lending platforms.
We apply machine learning and interpretability tools to large-scale lending data to identify the strongest predictors of loan default. Our goal is to improve credit assessment, inform smarter lending decisions, and enhance investor confidence in the evolving alternative finance ecosystem.
We are on a mission to:
- Build an open, data-driven framework for assessing default risk in P2P lending.
- Identify borrower and loan traits most predictive of default.
- Apply interpretable machine learning to improve credit assessment.
- Support investors, platforms, and regulators with transparent risk insights.
-
P2P platforms offer flexible credit access to millions, yet face a persistent challenge: borrower default. Unpaid loans hurt investors, threaten platform stability, and erode trust in digital finance.
-
Conventional credit scoring may miss key behavioral signals. Many borrowers, especially younger users, accumulate invisible debt across platforms. Without reliable models, lenders can't detect risk early or fairly.
-
By studying a large dataset of loan records and repayment history, we aim to reveal the hidden indicators of credit default risk and build interpretable models for real-world risk prediction.
What are the key borrower and loan characteristics that best predict default risk in peer-to-peer (P2P) lending platforms in the United States?
Our aim is to uncover the key drivers of default risk in P2P lending systems. Using historical data from Lending Club, we analyze borrower behavior, loan characteristics, and repayment outcomes to predict risk.
We train models that balance accuracy with explainability, enabling decisions that are both data-backed and transparent. The ultimate goal is to build tools that help platforms and investors reduce risk and improve outcomes.
Our project explores the intersection of behavioral finance and machine learning, with a focus on peer-to-peer (P2P) credit risk prediction. We aim to:
- Identify key borrower and loan features linked to default outcomes
- Build predictive models using Lending Club loan performance data
- Analyze behavioral and demographic traits influencing credit risk
- Apply feature importance tools to surface critical default indicators
- Support fairer, data-driven credit assessment in alternative lending
All datasets are stored in our
/1_datasets/
folder. Cleaning and preparation scripts are in
/2_data_preparation/.
The primary dataset used in our analysis is the Lending Club loan dataset, which includes over 2 million loans with borrower traits and repayment outcomes.
Key Features:
- Borrower: employment length, income, FICO score
- Loan: amount, term, purpose, interest rate
- Credit history: earliest credit line, open accounts, delinquencies
- Outcome: loan status (fully paid, charged-off)
To address our research question, we analyze Lending Club data to identify the borrower and loan features that best predict default risk.
Our modeling approach includes:
-
Data Cleaning: Filter loans with known outcomes, remove anomalies, and handle missing values for consistent analysis.
-
Feature Engineering: Create meaningful variables from raw data, such as debt-to-income ratios, credit history flags, and loan grade scores.
-
Exploratory Analysis: Visualize patterns of default by borrower demographics, loan purpose, FICO ranges, and installment size.
-
Modeling Techniques: Use classification models like
Logistic Regression, Random Forest, and XGBoost to estimate default likelihood. -
Interpretability Tools: Apply SHAP analysis and feature importance methods to explain model decisions and highlight key predictors.
-
Validation: Evaluate models with train-test splits and performance metrics (AUC, accuracy, recall) to ensure generalizability and robustness.
Our analysis of the Lending Club dataset reveals significant predictors of loan default. By employing a suite of machine learning models, we've identified key financial and behavioral traits that signal heightened credit risk.
We trained and evaluated three classification models to predict loan default. The models were optimized to handle class imbalance, ensuring that the minority class (defaulted loans) was given appropriate weight. XGBoost emerged as the top-performing model, demonstrating the best balance of precision and recall.
| Model | ROC | Prec (Def) | Rec (Def) | F1 (Def) |
|---|---|---|---|---|
| XGBoost | .72 | 0.32 | 0.67 | 0.44 |
| Logistic Regression | .71 | 0.31 | 0.67 | 0.43 |
| Random Forest | .71 | 0.55 | 0.06 | 0.10 |
Performance metrics are reported on the test set.
Feature importance analysis using both Random Forest and XGBoost, complemented by SHAP (SHapley Additive exPlanations) values from the XGBoost model, highlighted several critical factors in predicting loan defaults. The most influential features include:
- Interest Rate (
int_rate): Higher interest rates are strongly correlated with a higher probability of default. This is often the most significant predictor. - Loan Grade and Sub-Grade: The assigned loan grade (A-G) by the platform is a powerful indicator of risk, with lower grades showing much higher default rates.
- FICO Score (
fico_score): As expected, lower FICO scores are a primary indicator of credit risk. - Debt-to-Income Ratio (
dti): Borrowers with a higher percentage of their income going towards debt payments are more likely to default. - Annual Income (
annual_inc): Lower annual income is associated with a higher risk of default. - Loan Amount (
loan_amnt): Larger loan amounts can represent a higher risk.
To better understand the model's decisions, we used SHAP summary plots. These visualizations show the impact of each feature on the prediction for individual loans. For example, a high interest rate pushes the prediction towards default, while a high FICO score pushes it towards repayment.
This provides a transparent view into our model, allowing for interpretable, data-driven lending decisions. Our findings can help investors and platforms better assess risk and improve outcomes in the P2P lending market.
The SHAP summary plot above shows the impact of the top features on the model's output. Each point represents a single loan from the test set. The color indicates the feature's value (red is high, blue is low), and the position on the x-axis shows the feature's impact on the default prediction.
To share our insights with regulators and researchers, we built the DataCents Default Risk Dashboard - an interactive web application powered by our XGBoost model trained on over 2 million loan records.
Key Features:
- Instant Risk Scoring: Enter loan details for immediate risk assessment
- Transparent Analysis: SHAP-based explanations of key factors
- Interactive Testing: Modify parameters to explore different scenarios
- User-Friendly Interface: Accessible design for all stakeholders
🎤 Final Presentation: View our presentation on Canva
📋 For detailed communication strategy documentation, visit /5_communication_strategy/
Our repository is organized into key sections:
/0_domain_study/- Financial domain research and background/1_datasets/- Financial datasets and market data/2_data_preparation/- Data cleaning and preprocessing scripts/3_data_exploration/- Initial data analysis and visualization/4_data_analysis/- Advanced analysis and modeling/5_communication_strategy/- How we share our findings/6_final_presentation/- Final project presentation
- The data is from 2007–2018, so a lot has changed since then (especially after COVID).
- There's still class imbalance — even with reweighting, predicting default is hard.
- Try time-series models to predict default earlier in the loan cycle.
- Test the model on more recent data or on another P2P platform.
- Study if the models show bias (e.g., by demographic).
-
Clone and setup
# Clone the repository git clone https://github.com/MIT-Emerging-Talent/ET6-CDSP-group-15-repo.git cd ET6-CDSP-group-15-repo # Create environment conda env create -f environment.yml conda activate datacents # Or install manually pip install -r requirements.txt
-
Start exploring
# Launch Jupyter Notebook jupyter notebook
Navigate to the 4_data_analysis directory to begin exploring our financial
data analysis.
We welcome contributions! Please see our CONTRIBUTING.md for guidelines.
This project is licensed under the MIT License - see the LICENSE file for details.
"The goal is to turn data into information, and information into insight." - Carly Fiorina
Join us as we make sense — and DataCents — out of information.