FinGPT Trader

A quantitative trading system that combines Falcon-7B language model-driven sentiment analysis with technical market analysis to identify cryptocurrency trading opportunities. The system's core innovation is its confidence-weighted sentiment scoring mechanism, which integrates both the sentiment value and the model's confidence level to generate more reliable trading signals.

System Overview

FinGPT Trader processes financial news and market data in parallel, combining both sources to identify potential trading opportunities. The system uses:

• Falcon-7B Model: Fine-tuned for financial sentiment analysis
• Confidence-Weighted Scoring: 60% confidence weight, 40% sentiment strength weight
• Technical Analysis: Standard technical indicators for market inefficiency detection
• RSS-first News Ingestion: Crypto-native publisher RSS feeds, with NewsAPI fallback
• Event-Driven Architecture: Asynchronous processing of market events
• Risk Management: Basic position tracking and exposure monitoring

Current Implementation Status

✅ = Implemented | 🚧 = Partially Implemented | 📅 = Planned | ⚠️ = Implementation Issues

Core Components

• Sentiment Analysis

- ✅ Falcon-7B financial sentiment model implementation - ✅ Confidence-weighted sentiment scoring - ✅ Adaptive correlation tracking between sentiment and price

• Market Data Processing

- 🚧 Price data collection from Binance (⚠️ API parameter issues) - ✅ Technical indicator calculation (RSI, MAs, volume analysis) - ✅ Market inefficiency detection - 🚧 Minimum order size handling (⚠️ Size calculations incorrect)

• Trading Execution

- ✅ Binance testnet integration - 🚧 Order execution (⚠️ Method parameter mismatches) - ✅ Position tracking - 📅 Tax-loss harvesting

• Risk Management

- 🚧 Basic position tracking - ✅ Simple exposure metrics - 📅 Value-at-Risk (VaR) calculations - 📅 Correlation-aware risk metrics

Technical Implementation

Confidence-Weighted Sentiment Analysis

The system's key innovation is its confidence-weighted sentiment scoring mechanism. Rather than using raw sentiment scores, FinGPT Trader weighs the model's confidence (60%) against sentiment strength (40%) to generate more reliable signals:

# Calculate confidence-weighted sentiment score
weightedscore = (confidence  0.6)  (abs(sentimentscore) * 0.4)
Generate signal if weighted score exceeds detection threshold
if weightedscore > detectionthreshold:
    signals.append({
        'symbol': pair,
        'type': 'SENTIMENT',
        'direction': 1 if sentiment_score > 0 else -1,
        'strength': abs(sentiment_score),
        'confidence': confidence,
        'price': current_price,
        'timestamp': datetime.now()
    })

This approach helps filter out low-confidence predictions, reducing false signals.

Market-Sentiment Correlation Tracking

The system maintains a history of sentiment and price movements to calculate correlation coefficients:

# Calculate correlation between sentiment changes and price changes
correlation = np.corrcoef(
    sentimentchanges[:minlength],
    pricechanges[:minlength]
)[0, 1]
Update correlation record for position sizing
self.market_correlation[symbol] = {
    'value': correlation,
    'updated_at': datetime.now(),
    'samples': min_length
}

This correlation data helps adjust signal strength based on historical sentiment-price relationships.

System Architecture

graph TD
    A[Falcon-7B Model] -->|Sentiment Analysis| B[Market Analysis Engine]
    B -->|Signal Generation| C[Trading Core]
    D[Binance API] -->|Market Data| B
    C -->|Orders| D
    E[Risk Monitor] -->|Position Limits| C
    F[Portfolio Tracker] -->|Positions| C
    G[Market Inefficiency Detector] -->|Trading Signals| B
    H[News Data Feed] -->|Real-time News| A
    
    style A fill:#d0f0c0
    style B fill:#d0f0c0
    style C fill:#f0e0a0
    style D fill:#d0f0c0
    style E fill:#f0e0a0
    style F fill:#d0f0c0
    style G fill:#d0f0c0
    style H fill:#d0f0c0
    
    classDef implemented fill:#d0f0c0;
    classDef partial fill:#f0e0a0;
    classDef planned fill:#f0d0d0;

Configuration

The system uses a hierarchical YAML-based configuration system:

• trading.yaml: Core trading parameters, thresholds, and position sizing
• strategies.yaml: Strategy-specific parameters and detection thresholds
• model.yaml: ML model configurations
• services.yaml: External service connectivity parameters

News Sources

News ingestion is RSS-first. Default feeds are configured in config/services.yaml:

• CoinDesk: https://www.coindesk.com/arc/outboundfeeds/rss/
• Cointelegraph: https://cointelegraph.com/rss
• Cointelegraph Bitcoin: https://cointelegraph.com/rss/tag/bitcoin
• Cointelegraph Ethereum: https://cointelegraph.com/rss/tag/ethereum

NewsAPI is used as a fallback when RSS does not return enough relevant items. CryptoPanic is optional and disabled by default because its v2 API is plan-gated and the developer route returned 404 in this environment.

Setup Requirements

Prerequisites

• Python 3.8+ with asyncio support
• Windows, macOS, or Linux
• 8GB+ RAM recommended (for model loading)
• CUDA-compatible GPU recommended but not required

Environment Setup

# Create virtual environment
python -m venv venv
source venv/bin/activate  # or activate on Windows
Install dependencies
pip install -r requirements.txt
Windows-specific: Set console to UTF-8 mode
chcp 65001  # If running in cmd.exe

Configuration

# Set up your .env file with required API keys:
BINANCEAPIKEY=yourkeyhere         # Obtain from Binance Testnet
BINANCEAPISECRET=yoursecrethere   # Obtain from Binance Testnet
NEWSAPIKEY=yourkeyhere            # Optional fallback for news feeds
HUGGINGFACETOKEN=yourtoken_here     # Optional for model access
Optional paid/provider-specific fallback:
CRYPTOPANICAPIKEY=
CRYPTOPANICAPIPLAN=developer

Running the System

# Basic run
python main.py
Run with minimal console output
python main.py -q
Run with verbose debugging output
python main.py -v

Current Challenges

Known Issues

• API Connectivity

- CryptoPanic is disabled by default because API access is plan-gated - NewsAPI fallback can return zero articles for narrow crypto queries - Error handling needs continued hardening for provider outages

• Sentiment Analysis

- Method discrepancies between analyze() and analyze_text() - Unicode encoding errors in Windows environments - Inconsistent LLM response formats causing parsing errors

• Trading Execution

- Position sizing calculations produce orders below minimum requirements - Parameter mismatch in execute_trade() method - Inconsistent signal formats between processing methods

Immediate Improvements Needed

• API Connectivity

- Remove unsupported parameters in AsyncClient.create() - Add error handling for API timeouts and failures

• Method Compatibility

- Fix signature mismatch in trading execution methods - Standardize signal format between components

• Position Sizing

- Add validation for exchange minimum order sizes - Implement dynamic lookup of exchange requirements

• Error Handling

- Add more robust exception handling for LLM responses - Improve logging for system monitoring

Known Issues

• API Connectivity

- CryptoPanic API access is plan-gated and optional - RSS feeds are provider-controlled and may occasionally change format or URLs - Error handling for external API failures needs continued improvement

• Sentiment Analysis

- Method implementation discrepancies between analyze() and analyze_text() - Unicode encoding errors in Windows environments - Inconsistent response formats from the LLM causing "Unknown format code 'f'" errors - Type conversion issues when formatting sentiment scores

• Trading Execution

- Position sizing too small for minimum order requirements - Parameter mismatch in RoboService.execute_trade() method - Inconsistent signal format between different processing methods

• System Stability

- NoneType errors when awaiting non-coroutines - Uncaught exceptions in service initialization - UI vs. logging conflicts

Immediate Fixes Needed

• Binance Client Initialization

- Remove unsupported base_url parameter in AsyncClient.create() - Use only supported parameters: apikey, apisecret, and testnet

• Trading Execution

- Fix signature mismatch in RoboService.execute_trade() method - Standardize signal format between different processing stages

• Position Sizing

- Add validation for minimum order sizes before execution attempts - Implement dynamic lookup of exchange minimum requirements

• Sentiment Analysis

- Add type conversion for sentiment scores before formatting - Add exception handling for inconsistent LLM responses

Development Roadmap

Current development priorities:

• [x] Fix AsyncIO event loop issues
• [x] Improve error handling in API calls
• [ ] Enhance sentiment analysis prompt engineering
• [ ] Fix minimum order size calculation
• [ ] Fix Binance API client initialization parameters
• [ ] Resolve execute_trade() parameter mismatch
• [ ] Implement basic backtesting framework

Future enhancements:

• Advanced position sizing with Kelly Criterion
• Multi-asset portfolio optimization
• Value-at-Risk (VaR) calculations
• Sentiment-adjusted position sizing
• Correlation-aware risk metrics

Warning

⚠️ IMPORTANT: This system is in early development stage and not production-ready:

• Use testnet only - not suitable for real trading
• Expect instability and potential errors
• Many features described are planned but not yet implemented
• System requires significant technical knowledge to run properly

License

MIT License - See LICENSE for details.

Acknowledgements

This project builds on the foundation of several important open-source projects and research:

• FinGPT Research: Core ideas on financial sentiment analysis and market prediction using large language models were inspired by research from the FinGPT paper.

• llama.cpp: Our inferencing pipeline leverages the llama.cpp framework for running language models with minimal computational resources.

• Binance API: Trading functionality is built on the Binance exchange API and Python client libraries.

• Python AsyncIO: The asynchronous architecture is powered by Python's AsyncIO framework.

Contributing

Contributions are welcome! Please fork the repository and submit a pull request with your changes. Ensure to follow the code style and include tests for new features. See CONTRIBUTING.md for more details.