Adaptive RL Ensemble Strategy
- Outcome:
- Adaptive trading strategy that reduces portfolio volatility and improves risk-adjusted returns using an ensemble of reinforcement learning agents.
- Tech Stack:
- Python, PyTorch, PPO, A2C, TD3, Git
- Key Constraints:
- Adaptive allocation across varying market regimes, effective diversification to minimize drawdown, rigorous backtesting against SPY benchmark.
Developed at Queen's AI Club (QMIND), this project trains an ensemble of Reinforcement Learning agents (PPO, A2C, TD3) to rebalance a portfolio of 9 ETFs on a daily basis. Each agent learns inside a custom Gymnasium trading environment built on top of Stable-Baselines3, then their decisions are combined into a single adaptive allocation.
Rather than trusting a single backtest, I stress-tested the strategy across market regimes and modeled realistic 10 bps transaction costs. The results held up under scrutiny:
- Risk-adjusted returns: an annualized Sharpe ratio of 1.14, achieved at roughly half the drawdown of a passive S&P 500 benchmark.
- Statistical confidence: performance validated with 10,000-sample bootstrap confidence intervals, so the edge wasn't just a lucky run.
- Adaptive diversification: the ensemble shifts weight across ETFs as conditions change, smoothing volatility instead of riding any single position.

