Inspired by deep reinforcement learning breakthroughs in sequential decision-making, this project applies high-frequency Proximal Policy Optimization (PPO) to quantitative trading.
Key Technical Contributions:
- Level-2 Order Book Simulation: Custom OpenAI Gym/Farama Gymnasium environment modeling tick-by-tick order book depth, latency slippage, and execution impact.
- Actor-Critic Neural Policy: Transformer-based feature extractor mapping microstructure imbalances to optimal buy/sell/hold policy logits.
- Risk-Adjusted Reward Shaping: Direct optimization of Sortino & Sharpe ratios penalized by maximum drawdown constraints.