This paper addresses the challenge of ensuring realism in limit order book (LOB) market simulations, which are critical for training and testing trading strategies, especially those using reinforcement learning. The authors survey the literature to compile a comprehensive set of realism metrics, or stylized facts, that characterize real market behavior, including properties of asset return distributions (e.g., heavy tails, volatility clustering), order flow (e.g., power-law order sizes, interarrival times), intraday patterns, and market impact. They provide mathematical formulations for these metrics and use them to evaluate two agent configurations within the ABIDES simulation platform: one composed solely of zero-intelligence (ZI) agents and another mixing ZI agents with minimal strategic agents (market maker, heuristic belief learning, and momentum agents).

Experimental results show that while both configurations capture some stylized facts (e.g., heavy tails and volatility clustering), the mixed configuration generally performs better, producing statistics closer to historical data from U.S. equities. However, significant discrepancies remain, particularly in reproducing the absence of return autocorrelations, realistic order interarrival times, and the U-shaped intraday volume pattern. The authors conclude that current agent-based simulations still fall short of full realism and suggest that incorporating online learning and adaptive agents, which can evolve their strategies in response to market conditions, is a promising direction for improving fidelity. This work serves as a benchmark for future research, providing a standardized set of metrics to measure progress in market simulation realism.