DeFi Credit Scoring System for Aave V2 Protocol

Overview

This project implements a comprehensive machine learning-based credit scoring system for DeFi wallets interacting with the Aave V2 protocol. The system analyzes transaction patterns, behavioral indicators, and risk management practices to assign credit scores between 0-1000, where higher scores indicate more reliable and responsible usage.

Architecture

Core Components

1. Data Processing Pipeline: Extracts and flattens nested JSON data from MongoDB exports
2. Feature Engineering Engine: Creates 15+ behavioral and financial features
3. ML Scoring Model: Random Forest regression
4. Risk Assessment Framework: Multi-dimensional risk evaluation
5. Analysis & Visualization: Comprehensive scoring analysis and insights

Processing Flow

Raw JSON Data → Data Extraction → Feature Engineering → Model Training → Score Calculation → Analysis & Insights

Feature Categories

1. Transaction Metrics

• Volume Indicators: Total volume, transaction counts
• Action Diversity: Unique actions, action ratios

2. Temporal Behavior

• Consistency: Activity duration
• Frequency: Transactions per day

3. Portfolio Management

• Diversification: Unique assets
• Balance: Deposit/borrow/repay/liquidation ratios

4. Risk Management

• Leverage Ratios: Borrow-to-deposit, utilization ratios
• Repayment Behavior: Repay-to-borrow ratios
• Liquidation Exposure: Liquidation frequency and ratios

Scoring Methodology

Base Model

• Algorithm: Random Forest Regression
• Target Variable: Composite risk score (higher is better):
  • risk_score = (liquidation_ratio * 2) + (1 - repay_ratio) + borrow_ratio
  • Model is trained to predict -risk_score (so higher model output = lower risk)
• Features: 15+ engineered features, standardized

Score Adjustments

• Positive Factors:
  • High repay ratio, low borrow/liquidation ratios
• Negative Factors:
  • High liquidation or borrow ratio, low repay ratio

Final Score Calculation

raw_prediction = model.predict(features)
normalized_score = (raw_prediction - min) / (max - min)
credit_score = normalized_score * 1000  # Scale to 0-1000

Usage

Prerequisites

pip install pandas numpy scikit-learn matplotlib seaborn joblib

Running the Scorer

# Prepare features_df as shown in the notebook
from sklearn.ensemble import RandomForestRegressor
from sklearn.preprocessing import MinMaxScaler

# ... (feature engineering code)

# Model training and scoring
scaler = MinMaxScaler()
X_scaled = scaler.fit_transform(X)
model = RandomForestRegressor(n_estimators=100, random_state=42)
model.fit(X_scaled, y)
preds = model.predict(X_scaled)
scores = 1000 * (preds - preds.min()) / (preds.max() - preds.min())
features_df['credit_score'] = scores.astype(int)

Output Files

• credit_score_model.pkl: Trained model
• cleaned_wallet_transactions.csv: Cleaned transaction data
• wallet_features_and_scores.csv: Wallet-level features and scores

Score Interpretation

Score Ranges

• 900-1000: Excellent credit, highly reliable wallets
• 700-899: Good credit, responsible usage patterns
• 500-699: Fair credit, moderate risk indicators
• 300-499: Poor credit, concerning behaviors
• 0-299: Very poor credit, high-risk or bot-like behavior

Key Indicators

High-Quality Wallets (800+)

• Consistent, long-term activity
• Strong repayment history
• Balanced leverage usage
• Human-like transaction patterns

Low-Quality Wallets (300-)

• Irregular or bot-like patterns
• High liquidation exposure
• Poor repayment behavior
• Excessive leverage usage

Model Validation

Performance Metrics

• Feature Importance: borrow_ratio, repay_ratio, liquidation_ratio are top drivers
• Score Distribution: Well spread, with clear separation between risky and responsible wallets
• Behavioral Consistency: Low-score wallets are riskier, high-score wallets are more responsible

Robustness Checks

• Feature importance and correlation analysis
• Behavioral analysis of low/high score wallets

Extensibility

• Add new features or risk factors easily
• Support for additional DeFi protocols
• Real-time scoring capabilities
• Advanced ML models (e.g., XGBoost)

Technical Considerations

• Optimized for 100K+ transaction datasets
• Memory-efficient feature engineering
• Robust error handling for missing data
• Feature normalization and scaling

Future Enhancements

1. Real-time Scoring: Live transaction monitoring
2. Cross-protocol Analysis: Multi-DeFi platform scoring
3. Predictive Modeling: Future risk prediction
4. Market Integration: Price and volatility considerations

Contributing

Contributions are welcome for new features, models, and optimizations.

License

Open source under MIT License.