Realism of closed-loop agent-based generation under feedback

Characterize the realism of learned agent-based market dynamics under closed-loop feedback between generated order flow and the limit order book, including whether the observed spread mismatch persists over longer horizons and under injected-agent scenarios for measuring market impact.

Background

The paper evaluates closed-loop generation by sampling messages from the learned model and executing them in the matching engine. Although all rollouts maintain a valid two-sided book for 2,048 steps, the agent-based family exhibits a substantial spread discrepancy relative to held-out episodes, indicating that small modeling errors may compound when generated order flow interacts with the evolving book.

The authors explicitly leave unresolved whether the learned agent dynamics remain realistic under feedback. They also identify longer-horizon rollouts and scenarios with an injected agent, such as experiments measuring market impact, as settings in which this unresolved realism question could be investigated.