StockMARL: multi-agent reinforcement learning for trading

What it is

StockMARL: A Novel Multi-Agent Reinforcement Learning System to Dynamically Improve Trading Strategies. A market populated by diverse reactive agents plus one DQN agent whose observation space is built from the behaviour of the others — it learns to trade against the crowd rather than against raw prices alone.

• Behaviour-driven observation space
• Automated per-episode trade-history exports for post-hoc analysis
• Evaluated on unseen test data against the rule-based agents

Results

• 12.23% yearly money-weighted return, 15.9% cumulative return across S&P 500 stocks
• Low trade volatility and the leading profitability score among all agents