A multi-agent game-theoretic adversarial training framework with latency-aware reward shaping for robust high-frequency trading strategy design

Abstract High-frequency trading now plays out at microsecond and even nanosecond granularity, and profitability is shaped jointly by adversarial market dynamics and by acute sensitivity to execution latency. Most existing frameworks, somewhat surprisingly, still treat these two pressures in isolation, leaving robustness and timing largely uncoupled. This study sets out to unify adversarial robustness training with latency-aware policy optimization inside a single strategy-design framework, hereafter referred to as MAGAT — a multi-agent game-theoretic adversarial trading system. Methodologically, MAGAT pits a Protagonist Agent — trained by multi-agent proximal policy optimization under centralized training with decentralized execution — against an Adversary Agent that searches for worst-case perturbations using gradient-free evolution strategies; the two roles alternate in a minimax loop whose fixed point is interpreted as an approximate, not certified, equilibrium, and the approximation is itself diagnosed by tracking the joint best-response gap. A Latency-Aware Reward Shaping (LARS) term penalizes aggressive orders in proportion to the logarithm of realized delay, while an FPGA, INT8-GPU, and kernel-bypass pipeline targets sub-700-nanosecond tick-to-order latency. Performance is assessed entirely through event-driven simulation: matching-engine replay over Level-3 LOBSTER limit order book data for three U.S. equities (AAPL, MSFT, INTC, 2022), with a configurable delay-injection platform reproducing uniform, Pareto, and bursty latency profiles. Across the four stress scenarios MAGAT sustains Sharpe ratios of 1.97–2.31 (95% bootstrap CI ± 0.06) and survival rates of 89–96% (± 2.1pp), against 1.18 and 69% for the strongest single-agent baseline; latency elasticity falls roughly fourfold (from 0.54 to 0.12) and the measured 99th-percentile execution latency stays near 683 nanoseconds. These results refer strictly to simulation and should not, on their own, be read as evidence of deployable live performance.