RESEARCH · GLOBAL
arXiv: Deep RL agents achieve supra-competitive outcomes in trade execution
Study examines whether deep reinforcement-learning agents in shared liquidation environments can coordinate to achieve better execution (lower implementation shortfall) than game-theoretic competitive benchmarks. Finds learned memory structures sustain collusion-like outcomes.
WHY IT MATTERS
Signals risk of emergent agent coordination in execution systems. If BFSI deploying multi-agent systems for trade/settlement, supervisors must monitor for unintended collusion. Custody and execution desks must audit RL agent interaction logs.
Source: arXiv · 2026-05-21