AI Hedge Fund

Advanced Algorithmic Trading System with PatchTST & PPO Reinforcement Learning

An enterprise-grade portfolio management and automated trading system leveraging state-of-the-art Deep Learning (PatchTST) and Reinforcement Learning (PPO) to navigate financial markets.

Key Features

1. State-of-the-art Forecasting

• PatchTST Model: Utilizes the latest Transformer-based architecture for time-series forecasting.
• Enhanced Accuracy: Significantly outperforms traditional models (LSTM, ARIMA) on long-sequence forecasting.
• Hyperparameter Tuning: Integrated Bayesian optimization for model fine-tuning.

2. Intelligent Trading Agent (PPO)

• Reinforcement Learning: Uses Proximal Policy Optimization (PPO) to make autonomous trading decisions.
• Enhanced Features: 14+ technical signals including MACD, Stochastic RSI, Bollinger Bands, ADX, and Volume Analysis.
• Dynamic Scoring: Normalizes technical indicators into a 0.0-1.0 scoring system for stable agent training.

3. Unified Validation System

• Aligned Modes: Predict and Backtest modes share the exact same feature engineering and agent logic.
• Comprehensive Backtesting: Validate strategies with historical data, transaction costs, and slippage.
• Performance Metrics: Automatic calculation of Sharpe Ratio, Max Drawdown, Win Rate, and Total Return.

4. Robust Engineering

• Modular Architecture: Clean separation of Data, Trading, and Modeling layers.
• Interactive CLI: Rich terminal interface with progress bars and real-time status updates.
• Multi-Asset Support: Ready for Stocks, Crypto, Forex, and Commodities.

Installation

# Clone repository
git clone https://github.com/cyclocerine/HedgeFund.git
cd HedgeFund

# Create virtual environment
python -m venv venv
# Windows:
venv\Scripts\activate
# Linux/Mac:
source venv/bin/activate

# Install dependencies
pip install -r requirements.txt

Usage Guide

The unified CLI (scripts/run_cli.py) is the main entry point for all operations.

1. Predict Future Prices & Generate Signals

Run the system on the latest data to get actionable insights.

# Basic Prediction
python scripts/run_cli.py --ticker BMRI.JK --mode predict

# With PPO Trading Signals & Tuning (Recommended)
python scripts/run_cli.py --ticker BMRI.JK --mode predict --tune --ppo --ppo-episodes 200 --forecast-days 30 --save-results

2. Backtest Trading Strategy

Validate the PPO agent's performance on historical data.

# Run PPO Backtest (200 Episodes)
python scripts/run_cli.py --ticker BMRI.JK --mode backtest --strategy PPO --tune --ppo-episodes 200 --initial-balance 100000000 --save-results

CLI Arguments

Argument	Description	Default
--ticker	Asset symbol (e.g., BMRI.JK, BBCA.JK)	Required
--mode	predict or backtest	Required
--strategy	Trading strategy (PPO, Trend Following, Mean Reversion)	PPO
--tune	Enable hyperparameter tuning	False
--ppo-episodes	Number of training episodes for PPO	200
--forecast-days	Days to forecast into the future	30
--initial-balance	Starting capital for backtest	100,000,000
--save-results	Save plots and CSVs to results/	False

System Architecture

+-------------------+       +----------------------+
|    Market Data    | ----> |     Data Pipeline    |
+-------------------+       +----------------------+
                                       |
                                       v
                            +----------------------+
                            |  Feature Engineering |
                            +----------------------+
                                       |
                  +--------------------+--------------------+
                  |                                         |
                  v                                         v
        +------------------+                      +------------------+
        |  PatchTST Model  |                      |    PPO Agent     |
        +------------------+                      +------------------+
                  |                                         |
                  v                                         v
        +------------------+                      +------------------+
        | Price Prediction |                      |  Trading Signals |
        +------------------+                      +------------------+
                  |                                         |
                  +--------------------+--------------------+
                                       |
                                       v
                            +----------------------+
                            |    Trading Engine    |
                            +----------------------+
                                       |
                            +----------------------+
                            |     Risk Manager     |
                            +----------------------+
                                       |
                                       v
                            +----------------------+
                            |  Portfolio Execution |
                            +----------------------+

Performance Verification

Latest comprehensive test results (Jan 2026):

Metric	Predict Mode	Backtest Mode
Ticker	BMRI.JK	BMRI.JK
Model	PatchTST (Tuned)	PatchTST (Tuned)
Agent	PPO (200 Eps)	PPO (200 Eps)
Best Reward	+4.10	+3.70
Portfolio	Stable Growth	+240% Return

Contributing

Contributions are welcome! Please examine the src/ directory for core logic:

• src/models/patchtst_model.py: Deep Learning Forecast Model.
• src/trading/ppo_agent.py: Reinforcement Learning Agent.
• src/data/feature_engineering.py: Technical Indicator Processing.

References

• Chan, E. P. (2013). Algorithmic Trading: Winning Strategies and Their Rationale
• Sutton, R. S., & Barto, A. G. (2018). Reinforcement Learning: An Introduction
• De Prado, M. L. (2018). Advances in Financial Machine Learning
• Murphy, J. J. Technical Analysis of the Financial Markets
• Nie, Y., et al. (2022). A Time Series is Worth 64 Words: Long-term Forecasting with Transformers (PatchTST)

License

MIT

---

AI Hedge Fund - Intelligent Algorithmic Trading for the Digital Era