Reproducing Stylized Facts in Artificial Stock Markets with Price-Data- Trained Neural Agents

Traditional agent-based models (ABMs) of financial markets rely on handcrafted behavioral rules, limiting their realism, robustness, and scalability. To overcome these constraints, we propose a fully data-driven ABM in which each agent’s decision logic is learned directly from historical price series through neural-network training. In this framework, behavioral heterogeneity emerges endogenously as a result of differences in training depth and data composition, quantified by a continuous measure we term Fit Quality. This metric defines a spectrum of agent “personalities” ranging from flexible novice to experienced traders. When deployed within a standard limit-order-book environment, these data-trained agents collectively reproduce the major stylized facts of real financial markets—including heavy-tailed returns, volatility clustering, leverage effects, and gain–loss asymmetry—without fragile manual calibration. Ablation experiments confirm that these emergent properties originate from the agents’ learned intelligence rather than from the market mechanics alone. By enabling the construction of realistic market dynamics directly from historical data, our approach establishes a flexible and empirically grounded baseline model. It provides a robust experimental platform for exploring how variations in agent composition shape market stability, volatility, and systemic risk.