Algorithmic Trading Project

Overview

This project is focused on building an advanced algorithmic trading framework leveraging Reinforcement Learning (RL). The aim is to develop a robust trading agent that can manage portfolios efficiently, utilizing multi-dimensional reward functions and advanced RL techniques like Multi-Agent RL. The project is currently under development, and components are being iteratively refined.

---

Repository Structure

|-- airflow_home/           # Airflow DAGs and pipelines (no recent updates)
|-- binance/                # Binance API-related utilities
|-- logs/                   # Logs generated during training and evaluation
|-- saved_models/           # Directory for storing trained models
|-- .env                    # Environment variables (e.g., API keys)
|-- .gitignore              # Git ignore file
|-- data_fetch_job.py       # Script to manage data-fetching tasks
|-- db.py                   # Database interaction utilities
|-- fetch_data.py           # Fetch market data from external sources
|-- performance.py          # Evaluate portfolio performance metrics
|-- Portfolio.py            # Portfolio management class
|-- Position.py             # Position management class
|-- README.md               # This file
|-- requirements.txt        # Python dependencies
|-- SequentialTradingEnv.py # Custom RL environment based on PettingZoo
|-- Train.py                # Training pipeline for RL models

---

Key Components

Data Layer

• Database: PostgreSQL database to store historical market data.
• Data Fetching (fetch_data.py): Uses ccxt to pull OHLCV data from Binance.
• dbt: Supports data transformations and feature engineering (e.g., RSI, Bollinger Bands).

Airflow Automation

Airflow is used to automate data ingestion and preprocessing tasks:

• Data Fetching: Fetches binance OHLCV market data for selected trading pairs.

• DAGs: Orchestrates periodic data updates and backfills.

• data_fetch_job.py: Handles the periodic scheduling and execution of data-fetching tasks.

• fetch_data.py: Fetches market data from Binance and other APIs.

• db.py: Interfaces with the database for storing and retrieving market data.

Trading Logic

• Portfolio.py:

  • Manages portfolio allocations, including cash reserves and position rebalancing.
  • Ensures that total allocations (across pairs and cash) sum to 1.

• Position.py:

  • Handles individual trading positions.
  • Tracks open and closed positions, as well as profit and loss calculations.

Performance Evaluation

• performance.py:
  • Computes key performance metrics such as Sharpe ratio, drawdowns, and portfolio returns.
  • Supports backtesting and benchmarking.

RL Environment

• SequentialTradingEnv.py:
  • Implements the custom RL environment using PettingZoo.
  • Features:
    • Specialist Agents: Provide trade recommendations based on predefined strategies or signals.
    • Meta-Agent: Allocates capital based on specialist input.
    • Reward Mechanism: Multi-dimensional rewards per pair.
    • Cash Reserve Management: Allows holding unallocated funds for future opportunities.

Training Pipeline

• Train.py:
  • Orchestrates the training process using RLlib.
  • Handles hyperparameter tuning and model checkpoints.

---

Modeling Overview

The project uses a multi-agent reinforcement learning (MARL) framework:

• Specialist Agents:
  • Focus on individual trading pairs.
  • Recommend buy/sell/hold actions based on specific metrics.
• Meta-Agent:
  • Allocates funds across pairs and cash.
  • Ensures total allocations remain balanced.

The system integrates a multi-dimensional reward function to evaluate trades across pairs, promoting long-term profitability and robust portfolio management.