Deep Reinforcement Learning Market Maker
Production-Grade HFT System with Order Book Dynamics
This system demonstrates cutting-edge Deep RL (PPO) for optimal market making with realistic order book simulation, market microstructure modeling, and comprehensive analytics. The model was trained on a GPU via Google Colab and deployed here for inference on CPU.
Pre-Trained Model Overview
The PPO agent was trained on a GPU using Google Colab Pro. Below are the training metrics and reward curve from the training session.
Training Configuration
- Total Timesteps: 1,500,000
- Training Time: 68.7 minutes
- Hardware: NVIDIA A100-SXM4-80GB
- Algorithm: PPO (Proximal Policy Optimization)
- Reward Function: Avellaneda-Stoikov inspired (PnL change + inventory penalty)
- Environment: Custom Market Making Gym (12-D obs, 2-D continuous action)
Evaluation Results (10 episodes)
| Metric | Value |
|---|---|
| Sharpe Ratio | 0.000 |
| Mean PnL | $0.00 |
| Final PnL | $0.00 |
| Max Drawdown | 0.00% ($0.00) |
| Fill Rate | 0.0% |
| Mean Spread | $0.0123 |
| Mean Inventory | 0 shares |
| Mean Episode Return | 0.00 |
Model Status: Loaded and ready for inference.
Deep RL Market Maker v1.0.0 | Built with PyTorch, Stable-Baselines3, Gradio | Author: Spencer Purdy
Demonstrates: Deep RL (PPO), HFT, Market Microstructure, Order Book Dynamics, Real-time Trading Simulation