Source-linked AI summary
Agentic Limit Order Books: Phase Transitions and Market Impact
Jan Rosenzweig
TL;DR
The paper asks whether standard market-microstructure assumptions hold when adaptive agents populate a limit order book, and uses a statistical-mechanics framework to study phase behavior and market impact. It reports distinct collapsed, continuous, and frozen regimes and presents agentic order books as a simulation medium that reproduces behaviors encountered in live trading.
Problem
It is unclear whether standard assumptions about order arrivals, liquidity provision, and price impact hold in LOBs populated by adaptive agents, whose interactions create nonlinear feedback.
Source: §1 · p. 1Method
Results
Agentic interactions generate collapsed, continuous, and frozen regimes, each with distinct price-discovery and impact profiles.
Source: §6 · p. 5Takeaways & Limitations
The authors report that agentic order books reproduce a number of realistic behaviors encountered in live trading, supporting their use to test, validate, and risk-assess strategies.
Source: §6 · p. 6The homogeneous agent population enables mapping phase space by n and d but is unrealistic as a representation of live trading LOBs.
Source: §5 · p. 5
Abstract
from arXiv · showhide
We investigate the systemic macroscopic dynamics emerging from Limit Order Books (LOBs) populated exclusively by autonomous reinforcement-learning agentic traders. By formalizing agent interactions within a microscopic order-matching engine, we examine two fundamental quantitative phenomena: equilibrium phase transitions in order flow regime shifts, and the structural dynamics of market impact. We show that agentic LOBs exhibit distinct phase boundaries separating orderly price discovery from hyper-volatile cascade states, governed by critical thresholds in the number of agents and observable market depth. Furthermore, we demonstrate that market impact under agentic liquidity provision deviates from classical square-root dynamics, exhibiting distinct dissipative, balanced, and non-dissipative regimes under non-linear feedback loops.
1 Introduction
Adaptive agents can interact through nonlinear feedback that makes aggregate LOB behavior difficult to infer from individual traders. The paper frames agentic order books as reactive simulation environments and studies their phase boundaries and market-impact dynamics.
1 Introduction
- Adaptive agents co-adapt through nonlinear feedback, leaving it unclear whether standard assumptions about order arrivals, liquidity provision, and price impact still hold.The passages also caution that these macro-level phenomena cannot be predicted from single-agent behavior alone.Source: §1 · p. 1
- An agentic order book provides a reactive simulation environment in which trader responses play out, unlike historical simulations that rely on impact models to predict reactions.The stated aim is to test trading strategies against those responses.Source: §1 · p. 1
- The paper presents a mathematical and computational framework for studying aggregate dynamics in agentic limit order books.Its two stated targets are phase transitions and market impact under dynamic liquidity provision.Source: §1 · p. 1
- The paper maps agent population and market-depth parameters to phase boundaries between stable price discovery, frozen books, and liquidity collapse.It also examines the transition from dissipative to non-dissipative impact regimes.Source: §1 · p. 1
2 Model Setup & Agent Topology
The model is a discrete-time continuous double auction with autonomous market, scalping, and liquidity-providing agents. Agents respond to book events and callbacks by choosing among specified order actions.
2 Model Setup & Agent Topology
- The discrete-time continuous double auction contains three agent classes: Market Agents, Scalpers, and Liquidity-Providing agents.The book state is represented by bid/ask prices and quantities, and agents react to updates, trades, fills, and timed callbacks.Source: §2 · pp. 1–2
- Liquidity providers observe the book within ±d ticks of the best quotes and seek trading PnL using limit orders, with market orders mainly for inventory unwinding at ±L.They have a fixed depth horizon d and quantity limit L.Source: §2 · p. 2
- Scalpers observe only the touch and next-to-touch levels, while Market Agents submit random market buys and sells at fixed frequency.Scalpers use d = 2; Market Agent order sizes are bounded by a maximum limit.Source: §2 · p. 2
- Agents choose among limit buys or sells, market buys or sells, and modify/cancel actions.The mid-price evolution is introduced as a standard relation based on best bid and ask prices.Source: §2 · p. 2
3 Phase Transitions in Agentic LOBs
The model varies liquidity-provider count and observable depth to study agentic LOB phase behavior. Simulations distinguish collapsed, continuous-liquidity, and frozen regimes, with phase boundaries that depend differently on provider count and depth.
3 Phase Transitions in Agentic LOBs
- The analysis fixes one Market Agent and scales Scalpers as max(3, ⌊n/3⌋) while varying liquidity-provider count n and their shared observable depth d.At least three Scalpers are required to consistently form the touch and provide reference prices to liquidity providers.Source: §3 · p. 2
- Each parameter configuration uses 10,000 Monte Carlo paths, each run for 100,000 timer callbacks, tracking mid-price volatility and book-collapse probability.Collapse is defined as the book emptying on at least one side before path completion.Source: §3 · p. 2
3.1 Regime Shifts and Critical Thresholds
- Varying agent count n and depth d produces sharp transitions among collapsed, continuous-liquidity, and frozen regimes.Continuous markets retain finite non-zero volatility and price discovery, whereas frozen books have volatility far below one tick and stifled price discovery.Source: §3.1 · p. 2
- For n < n_c, insufficient liquidity provision leads to book extinction in finite time, with collapse probability p > 90%.The passage relates these abrupt shifts to critical power-law fluctuation behaviors found across financial markets.Source: §3.1 · p. 2
- Figure 1 contrasts solid, liquid, and vapour order books, with bids and asks shown as positive and negative quantities and the mid-price in black.The caption identifies the three phases from left to right.Source: Figure 1 · p. 3
- The collapsed-to-continuous boundary is fixed in n and invariant to depth d, while the continuous-to-frozen boundary scales linearly with agent density.Adding liquidity providers requires proportional expansion of spatial depth to maintain queue flexibility; once minimum capacity n_c is met, depth is secondary to structural stability under zero-mean non-directional flow.Source: §3.1 · p. 3
- The reported phase behavior assumes drift-free aggressive flow; significant directional bias eventually collapses the book even when n > n_c.The passage states that biased flow can fill a finite number of liquidity providers and cause collapse in finite time.Source: §3.1 · p. 3
4 Market Impact Dynamics
Market impact in the Continuous Phase ranges from dissipating to cascading as volatility crosses a critical boundary, while the Frozen Phase shows cracking or widespread unfreezing under an aggressive order.
4 Market Impact Dynamics
- The experiment injects a size-Q aggressive buy after 10 timer callbacks and measures impact by z-scoring paths with Q = 200 against Q = 0 controls.The z-score compares ensemble means and standard deviations of impacted and non-impacted paths over time.Source: §4 · p. 3
4.1 Market Impact in the Continuous Liquidity Phase
- In the Continuous Phase, impact unfolds through primary price displacement, secondary inventory-risk propagation, and subsequent decay or permanent shift.The primary stage consumes physical liquidity; scalpers absorb inventory imbalances in the secondary stage and pass residual risk to liquidity providers.Source: §4.1 · p. 4
- At σ ⪅4.21 ticks, impact decays; at σ ≈4.21 ticks it is permanent, while σ ⪆4.21 ticks produces self-sustaining cascades.The critical boundary separates dissipative, balanced, and non-dissipative responses.Source: Figure 3 · p. 4, §4.1 · p. 4
4.2 Market Impact in the Frozen Phase
- Below σ ⪅0.1 ticks, the Frozen Phase cracks a small subset of paths; above σ ⪆0.1 ticks, impact rapidly unfreezes queues across paths.The cracking regime leaves most paths locked, whereas the melting regime destabilizes queue positions.Source: §4.2 · p. 4, Figure 4 · p. 5
5 Discussion & Implications
The discussion links agentic feedback and phase shifts to execution and systemic-risk concerns, while noting that homogeneous agents are an unrealistic representation of live order books.
Agentic feedback and scope
- Even minimal homogeneous multi-agent LOBs produce non-equilibrium behavior, regime shifts, and phase transitions through feedback loops absent from classical rule-based models.The population homogeneity helps map phase behavior but is unrealistic for live trading LOBs.Source: §5 · p. 5
Execution and systemic risk
- Non-dissipative agentic regimes invalidate traditional square-root impact assumptions, and execution strategies must assess local phase state (n, d) to avoid triggering cascades.The paper also highlights adaptive circuit breakers in response to the sharp continuous-liquidity-to-collapse boundary.Source: §5 · p. 5
Risk management
- Transitions into frozen or collapsed phases invalidate liquidity adjustments assumed by standard VaR models, motivating stress tests around multi-agent phase shifts.The discussion notes that standard VaR models assume continuous market clearing.Source: §5 · p. 5
6 Conclusion
The paper presents a statistical-mechanics framework for agentic LOB regimes and argues that phase behavior matters for execution strategies. It also reports that A-LOBs reproduce realistic live-trading behaviors as a strategy-testing medium.
6 Conclusion
- The framework maps Agentic LOB phase transitions, liquidity-collapse boundaries, and non-dissipative market-impact functions into distinct Collapsed, Continuous, and Frozen regimes.The regimes have different price-discovery and impact profiles.Source: §6 · p. 5
- Execution strategies should account for LOB phase behavior because market responses can differ markedly across phases.Source: §6 · p. 5
- A-LOBs serve as a strategy simulation medium for testing, validation, and risk assessment, and reproduce a number of behaviors encountered in live trading.Source: §6 · p. 6