Optimization of High Frequency Quantitative Trading Strategies Based on Multi Agent Reinforcement Learning and Market Microstructure Awareness

The increasing complexity of electronic financial markets, characterized by fragmented liquidity, high velocity order flow, and strategic interactions among heterogeneous trading agents, poses significant challenges for traditional quantitative trading strategies. This study investigates the optimization of high frequency trading strategies through a multi agent reinforcement learning framework integrated with market microstructure awareness. Unlike single agent approaches that treat market conditions as exogenous, our framework models distinct trading agents such as liquidity takers, market makers, and arbitrageurs whose endogenous interactions shape price formation and execution dynamics within a limit order book environment. The proposed methodology incorporates granular order book features, including bid ask spreads, queue imbalances, and order flow intensity, alongside dynamic market impact costs and adaptive risk control mechanisms. Experimental evaluations are conducted using publicly available limit order book data from the LOBSTER database, covering high frequency trading activity across multiple NASDAQ stocks. The multi agent reinforcement learning framework is benchmarked against established execution strategies including Time Weighted Average Price and Volume Weighted Average Price. Performance metrics encompass execution shortfall, slippage costs, Sharpe ratio, and maximum drawdown. Findings indicate that the multi agent framework achieves superior adaptability to volatile market conditions and reduces adverse selection risks compared to single agent alternatives. This research contributes to the intersection of computational intelligence and financial economics, offering a reproducible foundation for high frequency strategy optimization in complex market microstructures.