Square-Root Price Impact Is Necessary for Endogenous Manipulation Cycles in Learning-Agent Markets

Battle Count: 84

Highly relevant for understanding endogenous market manipulation dynamics, the role of square-root price impact in enabling self-sustained oscillations, and the optimal-control structure of predatory strategies. Directly informs algorithmic execution (avoiding manipulation cycles), market microstructure modeling, and regulatory surveillance. The finding that cycles persist without retail herding (beta=0) and are architecture-independent suggests robust dynamical mechanisms applicable to real markets. The bang-bang optimal control interpretation provides actionable structure for strategy design.

Requires implementing: (1) agent-based market simulator with 20,000 retail agents, square-root impact, A-share mechanisms; (2) CMA-ES evolutionary optimization loop (800 generations, population 40, GPU-accelerated); (3) LSTM policy network (530 params); (4) mean-field ODE/SDE solver; (5) bifurcation analysis via binary search in parameter space; (6) entropy estimation (LZ76, Markov, permutation). The core simulation is moderate complexity, but the full pipeline with bifurcation mapping and multi-estimator entropy analysis is substantial. Total training ~4.5 hours on V100 GPU.

Full model parameters, architecture details, training procedure (CMA-ES settings, 800 generations, population 40), environment parameters, and ablation configurations are specified. Code and trained controllers promised upon publication. Supplemental material provides extensive detail. However, no GitHub repository is currently linked, and some parameters (herding scale HS, noise floor epsilon) require estimation from ABM statistics.