Multi-Modal Reinforcement Learning Trading Agent

This repository contains the full experimental artefact for an MSc Data Science dissertation at
Kingston University London.

The project investigates whether multi-modal data (price, news attention, macroeconomic indicators, and CNN-derived price embeddings) improves the performance and stability of a PPO-based reinforcement learning trading agent across cryptocurrency (BTCUSD) and foreign exchange (EURUSD) markets.

---

📌 Project Scope

• Offline, simulation-only reinforcement learning
• No live trading or real-money execution
• Historical backtesting using 30-minute OHLCV data
• Controlled ablation studies, hyperparameter tuning, and robustness evaluation
• Cross-asset transfer from BTCUSD to EURUSD

---

📂 Repository Structure

A detailed overview of the project layout is available here:
👉 Project Structure Overview

---

📄 Documentation (docs/)

All supporting documentation for ethics, licensing, provenance, and reproducibility is located in:

docs/ ├── project_summary.md # High-level description of the artefact and experiments ├── project_setup_checklist.md # Reproducibility and setup checklist ├── ETHICS.md # Ethics and compliance statement ├── DATA_LICENSE.md # Data source usage and licensing └── LICENSES.md # Third-party software licenses

---

🧠 Method Overview

1. Data Sources

   • Binance (BTCUSDT OHLCV)
   • OANDA (EURUSD OHLCV)
   • GDELT V2 GKG (global news event metadata)
   • Macroeconomic indicators (FRED, Bloomberg – institutional access)

2. Feature Engineering

   • Price-derived indicators
   • News attention and thematic signals
   • Macroeconomic context
   • CNN-based price embeddings and short-horizon direction probability

3. State Construction

   • Unified multi-modal state
   • Exported as Parquet, then converted to NumPy arrays
   • NumPy states consumed directly by the RL environment

4. Reinforcement Learning

   • Proximal Policy Optimisation (PPO)
   • Baseline training and feature ablations (BTC & EUR)
   • Hyperparameter tuning on BTC only
   • Cross-asset robustness testing on EUR
   • Buy-and-Hold benchmark comparison

---

📊 Results

• Evaluation metrics (return, Sharpe ratio, drawdown, trading activity) are stored per model in: results/<model_name>/

• Figures and plots used in the dissertation are stored in: visualisation/

---

⚖️ Ethics and Compliance

• No human subjects
• No personal or sensitive data
• No live trading
• No financial advice
• All data used under public or institutional academic licenses

See docs/ETHICS.md and docs/DATA_LICENSE.md for full details.

---

🔁 Reproducibility

• Environment specification is provided in environment.yml
• Code, data processing, and evaluation pipelines are modularised
• Final artefact is reproducible from raw data ingestion to result generation

---

🎓 Academic Context

This repository accompanies an experiment-based MSc dissertation submitted in partial fulfilment of the requirements for the MSc Data Science programme at Kingston University London.

---

Author: Gbemileke Micah
Student ID: K2457391