Queue-Reactive (QR) Model Overview

• Queue-Reactive models are stochastic systems where event intensities adjust based on local queue states, enabling nonlinear dynamics and endogenous feedback.
• They incorporate state-dependent Poisson processes and spatial interactions to model stability thresholds, phase transitions, and workload condensation.
• Applications include limit order book simulation, market microstructure analysis, and execution strategy testing using empirically calibrated likelihood methods.

The Queue-Reactive (QR) model refers to a class of stochastic models wherein the instantaneous (queue-specific) dynamics—arrival, service, or jump intensities—depend explicitly on the state of the queue or its neighborhood. Queue-reactive models were introduced independently in queueing theory for interacting spatial networks and, crucially, as a mathematical framework for state-dependent dynamics in limit order books (LOBs) in financial markets. The QR model family is unified by the principle that system evolution is dynamically modulated by the local workload, liquidity, or order state, thereby generating endogenous feedbacks, nonlinear stabilization, and critical phenomena.

1. Canonical Definitions and Mathematical Structure

Queue-reactive models are specified by state spaces comprising collections of local queues, each with integer-valued workload or volume. For each node $i$ (e.g., on $\mathbb{Z}^d$ spatial lattice, traffic ring, or LOB price axis), the local state is $X_i(t)\in\mathbb{N}$, with event intensities (arrivals, departures) that functionally depend on $X_i$ and potentially its spatial or market neighborhood.

For spatially interacting queues, a prototypical QR system is defined as follows (Sankararaman et al., 2017):

• Arrivals: Independent Poisson process at each $i$ with rate $\lambda$.
• Service Rate: For queue $i$, departure rate is $$\mu_i(X(t)) = \frac{X_i(t)}{\sum_j a_j\,X_{i-j}(t)}$$, where $\{a_j\}$ are interference weights with finite support.
• Global State: $X(t) = (X_i(t))_{i\in\mathbb{Z}^d}$.
• Processor Sharing: Individual customer service rates are modulated by the neighborhood workload via the denominator.

In LOB microstructure, QR models instantiate each price level as an independent or weakly coupled birth–death queue, generalizing classical zero-intelligence by making intensities $\mathbb{Z}^d$0 explicitly state-dependent on queue sizes (Huang et al., 2013, Bodor et al., 2024, Bodor et al., 15 Jan 2025). Continuous-time Markov chain dynamics follow, with state-dependent transition rates.

2. Theoretical Foundations and Stationary Regimes

The QR paradigm enables analytical characterizations of ergodicity, invariant measures, and pathwise construction under varying network and dependence architectures.

Spatially Interacting QR Networks

For $\mathbb{Z}^d$1 QR systems (Sankararaman et al., 2017):

• Stability Threshold: There exists a sharp phase transition for stationarity:

$\mathbb{Z}^d$2

The system exhibits a unique translation-invariant stationary regime with finite mean if and only if $\mathbb{Z}^d$3.

• Stationary Mean Formula:

$\mathbb{Z}^d$4

• Coupling from the Past: Minimal stationary regime is constructed via pathwise monotonicity, allowing sample-path minimality a.s.
• Sensitivity to Initialization: Uniformly bounded initial states are globally attracted to the minimal regime if stable; there exist large, sparse initial profiles for which divergence occurs even within the stable regime, indicating sensitivity to initial conditions.

LOB/Queue Microstructure QR

For Markov-queue LOBs (Huang et al., 2013, Bodor et al., 2024):

• Ergodicity: Each queue satisfies V-uniform ergodicity if there is negative drift for large queue sizes and bounded total birth rate.
• Invariant Law: For single queue with birth–death rates $\mathbb{Z}^d$5, the invariant law is

$\mathbb{Z}^d$6

• Multi-Queue Factorization: In the independent-queue QR, the stationary distribution factorizes over queues.

3. Model Extensions, Couplings, and Hybridizations

Queue-Reactive + Hawkes (QRH) Models

Hybridization with mutually-exciting Hawkes processes produces QRH models (Wu et al., 2019):

• QRH-I: Hawkes kernels $\mathbb{Z}^d$7 augment baseline QR rates at each level; event intensities depend both on current queue state and past event history.
• QRH-II: Full eight-process excitation at best bid/ask, with state-factor $\mathbb{Z}^d$8 modulating intensity; captures both state-dependence and order-flow endogeneity. Empirically, QRH models are strongly favored over pure QR or pure Hawkes by likelihood and goodness-of-fit tests.

Deep Learning MDQR and Endogenous/Exogenous Dynamics

Modern extensions replace parametric state dependence with deep neural networks, capturing high-dimensional cross-queue and market feature interactions (Bodor et al., 15 Jan 2025). State-space is enriched (e.g., with last-event types, spreads, trade imbalances), and cross-queue dependencies are learned, reproducing empirically observed microstructure stylized facts and market impact laws.

Multi-Asset and Efficient Price Coupling

Recent QR generalizations link multiple LOBs via a hidden Brownian efficient-price process $\mathbb{Z}^d$9 (Sfendourakis, 13 Jun 2025). Each asset's queue dynamics depend on $X_i(t)\in\mathbb{N}$0 (efficient-mid price gap), and macroscopic cross-asset correlations emerge exclusively via shared $X_i(t)\in\mathbb{N}$1.

Order-Size Awareness

Size-aware QR (SAQR) models make both event intensities and order-size distributions endogenous to the current queue state, leading to dramatically improved simulation of volatility, tails, and depletion events (Bodor et al., 2024).

4. Empirical Calibration and Practical Implementation

All QR models admit likelihood-based estimation from high-frequency event data. For the independent queue case:

• MLE Formula:

$X_i(t)\in\mathbb{N}$2

where $X_i(t)\in\mathbb{N}$3 is the waiting time in state $X_i(t)\in\mathbb{N}$4 before event $X_i(t)\in\mathbb{N}$5.

For marked (size-aware) or neural QR, event counts are binned in $X_i(t)\in\mathbb{N}$6 or higher-dimensional feature space, and trained via cross-entropy or negative log-likelihood.

Reference price and full-book re-initialization mechanisms allow endogenization of price moves and exogenous shocks, with calibration parameters $X_i(t)\in\mathbb{N}$7 fitted to return-volatility and mean-reversion statistics (Mariotti et al., 2022, Huang et al., 2013).

5. Macroscopic Phenomena: Stationarity, Phase Transition, and Condensation

QR models naturally give rise to phase transitions and collective effects:

• Phase Transition/Stability: Sharp threshold for queue-stationarity and workload divergence (Sankararaman et al., 2017).
• Condensation: In exclusion-based QR (multi-speed traffic), a power-law tail in the single-queue distribution can yield macroscopic condensation—a single queue absorbing a macroscopic fraction of the total workload once density exceeds a critical value (Furtlehner et al., 2011).
• Diffusive Limits: In multi-asset QR driven by efficient price, diffusive scaling yields observed price processes converging to correlated Brownian motion (Sfendourakis, 13 Jun 2025).

6. Applications in Limit Order Book Simulation, Execution, and Market Microstructure

QR models underpin contemporary LOB simulators, enabling:

• Execution Tactics Evaluation: QR-based simulators produce realistic price paths, queue statistics, and market impact profiles for backtesting market making or optimal execution strategies (Espana et al., 19 Nov 2025, Huang et al., 2013).
• High-Frequency Volatility Estimation: QR generates microstructure-conformant price series, essential for robust estimation and evaluation of volatility estimators (Mariotti et al., 2022).
• Reinforcement Learning: Embedding QR simulators in RL environments provides data-driven feedback with endogenous impact, facilitating training of meta-order execution and market-making algorithms without requiring closed-form impact models (Espana et al., 19 Nov 2025, Bodor et al., 15 Jan 2025).

7. Significance, Validation, and Comparison

Empirical validations show QR-based models reproduce key market stylized facts—queue distributions, cross-queue correlations, spread dynamics, volatility clustering, and impact laws—whereas simplistic zero-intelligence or pure Hawkes models are strongly rejected. Hybrid and deep extensions retain tractability and inference efficiency while capturing higher-order dependencies and providing a solid basis for both theoretical analysis and practical applications (Wu et al., 2019, Bodor et al., 15 Jan 2025, Bodor et al., 2024).

A plausible implication is that further integration of market signals, agent behaviors, and cross-asset dependencies into the QR formalism is likely to provide a unifying, analytically tractable, and empirically robust foundation for both microstructure modeling and financial simulation.