Deep Reinforcement Learning for Portfolio Allocation
Reproducing J.P. Morgan AI research to compare PPO against classical mean-variance optimization under market volatility.
- Reproduced a J.P. Morgan AI research setup to test whether a learning-based portfolio manager can allocate across market sectors as well as a classical optimizer.
- On ten held-out years of S&P 500 sector data (2012–2021), the agent matched the optimizer on Sharpe ratio (0.96 vs 0.94) and cut the worst drawdown (−15.4% vs −36.3%).
- Built a Gymnasium market-replay environment over those sector series, with softmax long-only weights and a Differential Sharpe reward so the agent is scored on risk-adjusted return.
- Trained a Stable-Baselines3 PPO policy and benchmarked each held-out year against a Ledoit–Wolf max-Sharpe mean-variance baseline in PyPortfolioOpt.
- Shipped a FastAPI backtesting API and a Vite/React dashboard for equity curves, rolling Sharpe, and allocation heatmaps.
- Python
- PyTorch
- Stable-Baselines3
- Gymnasium
- PyPortfolioOpt
- FastAPI
- React
- Vite