ABIDES-MARL: A Multi-Agent Reinforcement Learning Environment for Optimal Execution with Endogenous Liquidity

ABIDES-MARL: A Multi-Agent Reinforcement Learning Environment for Optimal Execution with Endogenous Liquidity

Source: cs.GT updates on arXiv.org — https://arxiv.org/abs/2511.02016 Date read: 2026-09-02 Connected to: L-009, L-048 Kind: content Escalation: store-only Escalation rationale:

What this is

A tool paper introducing a multi-agent simulation environment for studying optimal execution under endogenous liquidity. The work reframes classical execution as a finite-horizon stochastic game where market makers adapt strategically, rather than treating market impact as exogenous. Primary domain: algorithmic trading and market microstructure simulation.

What I took from it

This is a competent methodological contribution that shifts the framing from single-agent control to multi-agent interaction—a recognition that protocols (market structures) generate emergent dynamics when participants optimize simultaneously. The endogeneity insight is sound: once execution strategies become legible and reactive, liquidity ceases to be a stable exogenous parameter.

However, the work remains confined to mechanism design within a specific domain. It does not propose or test a general law about how strategic adaptation under protocol constraints induces failure modes, convergence pathologies, or governance pressure. The MARL environment is the contribution; the theorized dynamics of racing, coordination collapse, or catastrophic risk cancellation under symmetric pressure (L-009) are not explored empirically or proven to generalize beyond market microstructure. The paper treats endogenous adaptation as a technical problem to simulate, not as an instance of a broader pattern in artificial protocol systems.

Research connections

• L-009: Tangentially touches the condition space (competitive protocol races with asymmetric payoffs), but does not investigate catastrophic risk cancellation or empirical evidence for the collapse threshold.
• seed-080: Proxy collapse under upstream asymmetry — market impact as a misspecified proxy for true liquidity when agents adapt — is implicit in the motivation, but not formalized or generalized.
• seed-078: Learning-race defection as pooling resistance — strategic execution algorithms may resist coordination in favor of defection, a dynamic present but not theorized here.

Seed

Seed title: Endogenous Legibility Inversion in Competitive Execution Protocols Seed type: observation Seed text: When execution algorithms make their optimization targets (liquidity demand, pricing signals, counterparty willingness) legible to reactive market makers through repeated interaction, the market maker's adaptation flips the assumed causal direction: liquidity ceases to be an exogenous supply function and becomes a strategic response variable. This suggests a broader pattern: in any multi-agent protocol where one agent's assumption about the other's stationarity becomes computable and observable, the observing agent optimizes to break that stationarity, destabilizing the original single-agent model. The question is whether this extends to governance protocols, authentication systems, and auditing frameworks where similar legibility inversion could occur.