|
|
---
|
|
|
library_name: stable-baselines3
|
|
|
tags:
|
|
|
- reinforcement-learning
|
|
|
- finance
|
|
|
- stock-trading
|
|
|
- deep-reinforcement-learning
|
|
|
- dqn
|
|
|
- ppo
|
|
|
- a2c
|
|
|
model-index:
|
|
|
- name: RL-Trading-Agents
|
|
|
results:
|
|
|
- task:
|
|
|
type: reinforcement-learning
|
|
|
name: Stock Trading
|
|
|
metrics:
|
|
|
- type: sharpe_ratio
|
|
|
value: Variable
|
|
|
- type: total_return
|
|
|
value: Variable
|
|
|
---
|
|
|
|
|
|
# 🤖 Multi-Agent Reinforcement Learning Trading System
|
|
|
|
|
|
This repository contains trained Deep Reinforcement Learning agents for automated stock trading. The agents were trained using `stable-baselines3` on a custom OpenAI Gym environment simulating the US Stock Market (AAPL, MSFT, GOOGL).
|
|
|
|
|
|
## 🧠 Models
|
|
|
|
|
|
The following algorithms were used:
|
|
|
1. **DQN (Deep Q-Network)**: Off-policy RL algorithm suitable for discrete action spaces.
|
|
|
2. **PPO (Proximal Policy Optimization)**: On-policy gradient method known for stability.
|
|
|
3. **A2C (Advantage Actor-Critic)**: Synchronous deterministic policy gradient method.
|
|
|
4. **Ensemble**: A meta-voter that takes the majority decision from the above three.
|
|
|
|
|
|
## 🏋️ Training Data
|
|
|
|
|
|
The models were trained on technical indicators derived from historical daily price data (2018-2024):
|
|
|
* **Returns**: Daily percentage change.
|
|
|
* **RSI (14)**: Relative Strength Index.
|
|
|
* **MACD**: Moving Average Convergence Divergence.
|
|
|
* **Bollinger Bands**: Volatility measure.
|
|
|
* **Volume Ratio**: Relative volume intensity.
|
|
|
* **Market Regime**: Bull/Bear trend classification.
|
|
|
|
|
|
## 🔗 Related Data
|
|
|
|
|
|
* **Dataset Repository**: [AdityaaXD/Multi-Agent_Reinforcement_Learning_Trading_System_Data](https://huggingface.co/AdityaaXD/Multi-Agent_Reinforcement_Learning_Trading_System_Data)
|
|
|
* **GitHub Repository**: [ADITYA-tp01/Multi-Agent-Reinforcement-Learning-Trading-System-Data](https://github.com/ADITYA-tp01/Multi-Agent-Reinforcement-Learning-Trading-System-Data)
|
|
|
|
|
|
|
|
|
## 🎮 Environment (`TradingEnv`)
|
|
|
|
|
|
* **Action Space**: Discrete(3) - `0: HOLD`, `1: BUY`, `2: SELL`.
|
|
|
* **Observation Space**: Box(10,) - Normalized technical features + portfolio state.
|
|
|
* **Reward**: Profit & Loss (PnL) minus transaction costs and drawdown penalties.
|
|
|
|
|
|
## 🚀 Usage
|
|
|
|
|
|
```python
|
|
|
import gymnasium as gym
|
|
|
from stable_baselines3 import PPO
|
|
|
|
|
|
# Load the environment (custom wrapper required)
|
|
|
# env = TradingEnv(df)
|
|
|
|
|
|
# Load model
|
|
|
model = PPO.load("ppo_AAPL.zip")
|
|
|
|
|
|
# Predict
|
|
|
action, _ = model.predict(obs, deterministic=True)
|
|
|
```
|
|
|
|
|
|
## 📈 Performance
|
|
|
|
|
|
Performance varies by ticker and market condition. See the generated `results/` CSVs for detailed Sharpe Ratios and Max Drawdown stats per agent.
|
|
|
|
|
|
## 🛠️ Credits
|
|
|
|
|
|
Developed by **Adityaraj Suman** as part of the Multi-Agent RL Trading System project.
|
|
|
|
