Quantitative Trading with Deep Reinforcement Learning

The Opportunity

QuantaFin Capital, a proprietary trading desk in Milan managing a €120M portfolio focused on European equity derivatives, sought to augment their human traders with an AI-driven execution system. The goal was not to replace traders but to handle the high-frequency, low-margin statistical arbitrage opportunities that humans physically cannot react to fast enough — sub-second price dislocations across correlated instruments.

System Architecture: The Trading Brain

1. Market Data Infrastructure

Ultra-low-latency market data feed processing via kernel-bypass networking (DPDK).
Normalized order book data from Borsa Italiana, Eurex, and Euronext into a unified internal format.
Tick-by-tick storage in QuestDB (a time-series database optimized for financial data) — 5+ billion rows per day.

2. Reinforcement Learning Agent

Custom Proximal Policy Optimization (PPO) agent built with Stable-Baselines3.
State space: 127-dimensional feature vector including order book imbalance, VWAP deviation, realized volatility (Parkinson estimator), cross-asset correlation matrix, and Greeks exposure.
Action space: Continuous — position sizing from -1 (max short) to +1 (max long) with discrete hold option.
Reward function: Risk-adjusted PnL (Sharpe-penalized) with transaction cost awareness and maximum drawdown penalties.
Trained on 7 years of tick data (~180 billion data points) using curriculum learning: simple single-asset → multi-asset → portfolio-level.

3. Risk Management Layer

Hard-coded risk limits that cannot be overridden by the AI agent:
- Max position size: 2% of NAV per instrument.
- Max portfolio VaR (99%, 1-day): €1.2M.
- Kill switch: automatic liquidation if daily PnL drops below -€500k.
Real-time Greeks monitoring (Delta, Gamma, Vega, Theta) with auto-hedging.

4. Execution Engine

Smart Order Router (SOR) that fragments large orders across venues to minimize market impact.
FIX 4.4 protocol connectivity to multiple execution venues.
Latency budget: < 500μs from signal to order submission.

5. Monitoring & Explainability

Custom Grafana dashboards showing real-time PnL attribution, agent confidence scores, and feature importance (SHAP values).
Every trading decision is logged with the full state vector and the agent's action probability distribution for post-trade analysis.
Daily automated reports comparing AI performance vs. human trader benchmarks.

Technology Stack

ML/RL: Python, PyTorch, Stable-Baselines3, Ray (distributed training)
Data: QuestDB, Apache Parquet, Redis (feature store)
Execution: C++ (order gateway), FIX protocol, DPDK
Infrastructure: Bare-metal servers (co-located at Equinix ML1), Docker
Monitoring: Grafana, Prometheus, custom Python dashboards

Performance (12-month live track record)

Annualized Sharpe Ratio: 2.8 (vs. 1.4 for the human-only desk).
Maximum Drawdown: -3.2% (vs. -8.7% human desk).
Win Rate: 62% of trades profitable.
Incremental Alpha: +€4.8M attributable to the RL agent.
Latency: p99 signal-to-order = 380μs.