A multi-agent game-theoretic adversarial training framework with latency-aware reward shaping for robust...

Area of Science:

• Quantitative Finance
• Algorithmic Trading
• Machine Learning

Background:

• High-frequency trading (HFT) operates at microsecond/nanosecond speeds, demanding both adversarial robustness and low execution latency.
• Existing HFT frameworks often address market dynamics and latency in isolation, leading to suboptimal performance.
• There is a need for integrated strategies that simultaneously optimize for robustness against adversarial behavior and sensitivity to execution delays.

Purpose of the Study:

• To develop and evaluate a unified strategy-design framework, MAGAT (Multi-Agent Game-theoretic Adversarial Trading), that integrates adversarial robustness training with latency-aware policy optimization.
• To assess the performance of MAGAT against a strong single-agent baseline under simulated market stress conditions.
• To quantify the impact of latency on trading strategy performance and the effectiveness of latency-aware reward shaping.

Main Methods:

• MAGAT employs a Protagonist Agent trained via multi-agent proximal policy optimization and an Adversary Agent using gradient-free evolution strategies in a minimax loop.
• A Latency-Aware Reward Shaping (LARS) term penalizes aggressive orders based on realized delay.
• Performance was evaluated using event-driven simulation on LOBSTER limit order book data, with a delay-injection platform simulating various latency profiles.

Main Results:

• MAGAT achieved Sharpe ratios of 1.97-2.31 and survival rates of 89-96% across four stress scenarios, significantly outperforming a baseline (Sharpe ratio 1.18, survival rate 69%).
• Latency elasticity decreased approximately fourfold (0.54 to 0.12) with MAGAT.
• The system maintained a 99th-percentile execution latency near 683 nanoseconds, targeting sub-700-nanosecond tick-to-order latency with specialized hardware and software pipelines.

Conclusions:

• The integrated MAGAT framework effectively unifies adversarial robustness and latency optimization in high-frequency trading.
• MAGAT demonstrates superior performance and reduced latency sensitivity compared to traditional single-agent approaches under simulated adversarial conditions.
• The findings highlight the importance of joint optimization for robust and efficient HFT strategy design, though results are simulation-based.