Risk Estimation Agent for Wallet Account on EVM Network

A comprehensive system for analyzing cryptocurrency wallet risk through machine learning models and real-time data collection.

Overview

This system provides tools for:

• Data Collection: Gathering real-time data from Uniswap, Aave, cryptocurrency markets, and news sources
• Risk Modeling: Training machine learning models to predict coin and wallet risk scores
• Wallet Analysis: Comprehensive risk assessment of cryptocurrency wallets
• DeFi Position Monitoring: Tracking and analyzing DeFi protocol positions

Problem Statement

This project aims to evaluate the risk score of blockchain wallets. It is relatively large and complex, as evaluating a wallet's risk score numerically is challenging and there is no definitive dataset for validation.

To address this, we consider multiple factors, including:

• Wallet balances
• Active positions in DeFi protocols
• Transaction history

Methodology

1. Component-Level Risk Evaluation

Risk scores for different components of the wallet are calculated using a hybrid approach, combining rule-based and AI-based methods.

DeFi Protocols (Uniswap v3, Aave)

• Metrics are derived from existing research and academic papers
• While machine learning could be applied, leveraging expert research is more practical and saves time

Individual Coins and Blockchain Networks

• Rule-based methods are applied using metrics such as:
  • Total market capitalization
  • Maximum supply
  • Current price
  • Price volatility
• Reference methodology: Exponential Finance Whitepaper
• Additional research could refine assumptions or help develop new models

2. AI-Based Enhancements

AI is suitable due to the mathematical complexity of DeFi risk assessment and the lack of labeled training data.

Two main approaches were implemented:

1. Rule-Based Dataset Generation (coinRiskModel)

   • Training data is generated from predefined rule-based evaluations for coins, chains, and wallets

2. Pairwise Comparison Guided by Experts (pairWiseModel)

   • Relative risk between two entities (coins, chains, wallets) is compared using expert judgment or LLM guidance
   • Effective because absolute risk scores are difficult to obtain, but expert comparisons are reliable

3. Defining Risk Weights via AI

AI is used to define risk weights applied when calculating the total risk score from individual component scores:

• The total risk score is expressed as a linear regression over individual component scores
• Coefficients (weights) are learned from training data generated via pairwise comparisons
• This ensures the final score reflects both the relative importance of each component and expert-informed assessments

4. Transaction-Level Risk Scoring

Transaction-level risk is evaluated using:

• Transaction amounts
• Transaction frequency
• Accounts interacted with

This provides a granular understanding of wallet risk at the operational level.

Core Functions

1. Data Collection Functions

update_data_sources(config, args)

Updates data sources based on command line arguments.

Parameters:

• config: Configuration dictionary loaded from YAML
• args: Parsed command line arguments

Returns:

• List of successfully updated data sources

Features:

• Handles Uniswap, Aave, cryptocurrency, and news data updates
• Error handling for each data source
• Progress reporting and status updates

2. Model Training Functions

train_coin_risk_model()

Trains the coin risk regression model.

Features:

• Imports and trains coin risk model from ai.coinRiskModel
• Error handling and status reporting
• Returns success/failure status

pair_wise_risk_model()

Trains the pairwise risk model for relative risk assessment.

Features:

• Loads coin data from JSON
• Trains pairwise comparison model
• Handles training errors gracefully

risk_weight_model()

Trains the risk weight model using pairwise learning to optimize risk component weights.

Features:

• Loads mock wallet risk component data from JSON
• Creates pairwise training data with extreme true weights
• Trains logistic regression model on pairwise comparisons
• Optimizes weights for 7 risk components using L1 regularization
• Handles training errors gracefully and provides weight comparison analysis

3. Wallet Analysis Functions

analyze_wallet_risk_safe(wallet_address, config)

Safely analyzes wallet risk with comprehensive error handling.

Parameters:

• wallet_address: Ethereum wallet address or "mock" for testing
• config: Configuration dictionary

Returns:

• Risk analysis results or None if error

Features:

• Supports real wallet addresses via Etherscan API
• Mock wallet support for testing
• Returns structured risk analysis

Data Sources

1. Uniswap Data (utils/collectors/uniswap.py)

• Purpose: Collects Uniswap pool data for liquidity analysis
• Data: Pool addresses, token pairs, liquidity, fees, volume
• Usage: Risk assessment of DeFi liquidity positions

2. Aave Data (utils/collectors/aave.py)

• Purpose: Gathers Aave lending market information
• Data: Market addresses, assets, APYs, risk parameters
• Usage: Lending protocol risk evaluation

3. Cryptocurrency Data (utils/collectors/coin.py)

• Purpose: Collects general cryptocurrency market data
• Data: Prices, market caps, volumes, price changes
• Usage: Market risk assessment and trend analysis

4. News Data (utils/collectors/news.py)

• Purpose: Gathers news sentiment data for cryptocurrencies
• Data: News articles, sentiment scores, keywords
• Usage: Sentiment-based risk analysis

5. Etherscan Data (utils/collectors/etherscan.py)

• Purpose: Retrieves blockchain transaction data
• Data: Transaction history, token transfers, contract interactions
• Usage: Wallet behavior analysis and risk scoring

AI Models

1. Coin Risk Model (ai/coin_risk_model.py)

• Type: Regression model for absolute risk scoring
• Features: Market cap, supply, price volatility, volume analysis
• Output: Numerical risk score (0-100)

2. Pairwise Risk Model (ai/pairwise_model.py)

• Type: Learning-to-rank model for relative risk assessment
• Features: Pairwise comparisons between coins
• Output: Relative risk rankings and scores

3. Risk Weight Model (ai/riskWeightModel.py)

• Type: Pairwise learning-to-rank for deriving component weights
• Goal: Learn optimal weights for wallet risk components using only pairwise preferences
• Data: Mock pairwise comparisons produced with an "extreme" ground-truth weight vector to create strong training signal
• Methods:
  • Standard Logistic Regression (method='logreg')
  • L1-regularized Logistic Regression (method='logreg_l1', solver=saga) to encourage larger weight shifts
• Outputs:
  • trained_risk_weights.json with normalized weights (sum to 1)
  • wallet_pairwise_training_meta.json including true_weights used to generate labels

Wallet Analysis

Risk Components

The wallet risk calculator evaluates multiple risk factors:

1. Frequency Risk (15%)

   • Daily transaction patterns
   • Transaction frequency distribution
   • Peak activity periods

2. Amount Risk (15%)

   • Transaction volume analysis
   • Balance-to-volume ratios
   • Large transaction patterns

3. Interaction Risk (15%)

   • Address diversity
   • Interaction patterns
   • New address ratios

4. Recent Activity Risk (10%)

   • Recent transaction analysis
   • Activity trends
   • Temporal patterns

5. Contract Risk (15%)

   • Smart contract interactions
   • Suspicious contract detection
   • Zero-value transaction analysis

6. Balance Risk (20%)

   • Portfolio composition
   • Token diversification
   • Value concentration

7. DeFi Position Risk (10%)

   • Protocol exposure
   • Position values
   • Risk level assessment

Usage

Command Line Interface

The main entry point is test.py which provides a comprehensive CLI:

# Update all data sources
python test.py --update-all

# Train risk models
python test.py --train-risk-weight-model

# Analyze specific wallet
python test.py --wallet 0x7a29aE65Bf25Dfb6e554BF0468a6c23ed99a8DC2

# Use mock wallet for testing
python test.py --wallet mock

# Update specific data sources
python test.py --update-uniswap --update-coins

Future Improvements

1. Identify Additional Risk Factors

   • Explore more factors influencing wallet risk
   • Estimate their contribution using hybrid methods
   • Apply Principal Component Analysis (PCA) to understand factor importance

2. Expand AI Applications

   • Experiment with different model types and architectures
   • Apply a hybrid approach combining rule-based and AI-based methods for specific use cases

3. Develop a Virtual Simulation Environment

   • Reconstruct historical blockchain data to create a realistic environment
   • Backtest risk models to validate performance
   • Implement reinforcement learning with Direct Preference Optimization (DPO) to iteratively improve model accuracy and robustness

References

• Exponential Finance Whitepaper