Memory-Induced Supra-Competitive Outcomes Between Deep Reinforcement Learning Agents in Optimal Trade Execution

Battle Count: 78

Highly relevant to quantitative trading. The paper directly addresses optimal trade execution, a core problem in institutional trading. It reveals that the information architecture given to execution algorithms (whether they see price history, own past actions, or only current state) can qualitatively change whether learned behavior is competitive or supra-competitive. This has direct implications for algorithm design, regulatory compliance (avoiding tacit collusion), and understanding emergent market dynamics from automated execution systems.

Moderate to high complexity. Requires implementing: (1) a discrete Almgren-Chriss simulator with two temporary-impact conventions, (2) analytical Nash benchmarks (Schied-Zhang and Cordoni-Lillo), (3) DDQN training with simultaneous moves, (4) a Transformer-based sequence encoder for history-aware agents, (5) integrated gradient attribution for diagnostics, and (6) ex-ante schedule learning with perturbation rollouts. The mathematical framework is well-defined but the full experimental pipeline is substantial.

The paper provides detailed hyperparameters (Table 1), environment specifications, and benchmark formulas. However, no code repository is mentioned. The Almgren-Chriss simulator, DDQN implementation, and Transformer architecture details are described but not fully specified in code. The discrete Nash benchmarks (Schied-Zhang and Cordoni-Lillo) are analytically defined. Reproducibility would require implementing the full simulator and training pipeline from scratch.