This paper presents a multi-agent model designed to simulate the dynamics of a double auction market at microscopic time scales by faithfully replicating the mechanics of a limit order book (LOB). The agents in the model are 'zero-intelligence' in the sense that they lack utility maximization; instead, their decisions are driven by a stochastic variable called 'market sentiment,' which combines public information (e.g., perceived volatility and liquidity) with private information modeled as Gaussian noise. The model incorporates realistic order submission, cancellation, and matching processes, including the generation of limit and market orders based on a log-normal distribution for price aggressiveness and volume.

Despite its simplicity, the model successfully reproduces several empirical stylized facts of high-frequency financial markets. Key results include negative autocorrelation in returns (bid-ask bounce), leptokurtic return distributions with fat tails, volatility clustering with a Hurst exponent of 0.61, and a nonlinear price impact function that becomes sublinear for large volumes. The average LOB shape exhibits a peak in liquidity near the mid-price, and the spread shows persistence and power-law tails. The model also captures the mean-reverting nature of volume imbalance between buy and sell orders.

The findings suggest that many features of short-term market dynamics can be explained without assuming rational agents, relying instead on simple stochastic decision rules and memory effects from moving averages of public information. The paper concludes that large price movements are more likely caused by temporary liquidity shortages in the LOB rather than by large trade volumes, and it outlines future work to incorporate herding behavior, news, and market-making strategies.