Source-linked AI summary
Simulating and analyzing order book data: The queue-reactive model
Weibing Huang, Charles-Albert Lehalle, Mathieu Rosenbaum
TL;DR
The paper addresses how to model full limit-order-book dynamics while retaining lower-frequency financial properties. It builds a queue-reactive Markov model with reference-price switching, establishes its limiting behavior, and shows that it supports short-term prediction and market simulation for complex trading strategies.
Problem
Understanding limit order book dynamics is important for regulation, liquidity provision, transaction-cost reduction, and explaining emergent price behavior.
Method
The paper splits time into constant-reference-price periods, models the order book as a state-dependent Markov queuing system, and adds stochastic reference-price switches.
Results
The framework provides short-term predictions and is shown to be a relevant market simulator for complex trading tactics involving market and limit orders.
Takeaways & Limitations
The model supports execution-probability and price-movement analysis as well as pre-trade transaction-cost analysis for complex strategies.
Takeaways & Limitations
Execution probabilities may be slightly overestimated because the model does not track order identifiers and therefore does not fully capture queue-priority effects.
Abstract
from arXiv · showhide
Through the analysis of a dataset of ultra high frequency order book updates, we introduce a model which accommodates the empirical properties of the full order book together with the stylized facts of lower frequency financial data. To do so, we split the time interval of interest into periods in which a well chosen reference price, typically the mid price, remains constant. Within these periods, we view the limit order book as a Markov queuing system. Indeed, we assume that the intensities of the order flows only depend on the current state of the order book. We establish the limiting behavior of this model and estimate its parameters from market data. Then, in order to design a relevant model for the whole period of interest, we use a stochastic mechanism that allows for switches from one period of constant reference price to another. Beyond enabling to reproduce accurately the behavior of market data, we show that our framework can be very useful for practitioners, notably as a market simulator or as a tool for the transaction cost analysis of complex trading algorithms.
1 LPMA, University Pierre et Marie Curie (Paris 6) 2 Kepler-Cheuvreux 3 Capital Fund Management
The paper concerns limit order books, market microstructure, high-frequency data, and queue-based modeling.
- The paper focuses on limit order books, high-frequency data, queuing models, and market microstructure.
1 Introduction
The paper studies how limit order books evolve and develops a queue-reactive framework that combines constant-reference-price queue dynamics with reference-price moves. It evaluates the framework through short-term predictions and market-simulation applications.
- Understanding limit order book dynamics matters for regulation, liquidity provision, transaction-cost reduction, and explaining price behavior.
- The framework splits time into periods of constant reference price and models order-book dynamics as state-dependent queueing behavior.
- The paper extends the framework with reference-price moves and an exogenous component because endogenous order-flow randomness typically produces less volatility than observed.
- The models support short-term predictions including passive-order execution probabilities and price-increase probabilities.
- The queue-reactive model is presented as a market simulator for analyzing complex trading tactics and transaction costs.
2 Dynamics of the LOB in a period of constant reference price
For periods with a constant reference price, the paper introduces three models for limit-order-book dynamics through a shared general framework.
- Three limit-order-book models are jointly introduced for periods during which the reference price remains constant.
2.1 General Framework
The general framework represents the limit order book as a multidimensional continuous-time Markov jump process with state-dependent queue flows. Under drift and bounded-arrival assumptions, the process is ergodic.
- The limit order book is represented by a 2K-dimensional vector centered on a reference price, with bid and ask queues indexed by distance from that center.
- The framework models limit orders, cancellations, and market orders at each limit, allowing different constant order sizes across limits.
- The queue state evolves as a continuous-time Markov jump process on N^2K, with unit-sized jumps and generator entries determined by state transitions.
- The asymptotic analysis relies on an invariant probability measure and convergence of transition probabilities between states.
- Ergodicity follows under negative individual drift for large queues and a bounded incoming flow that prevents system explosion.
2.2 Data description and estimation of the reference price
The study uses Euronext Paris order-book data for two large-tick stocks and constructs a reference price to center the observed book, typically using the midprice when appropriate.
- Data description: The dataset records prices, volumes, and order counts through the fifth best limit on both sides whenever the order book changes.Observations from the first and last trading hours are removed because opening and closing auctions have distinctive features.
- Data description: The empirical study covers France Telecom and Alcatel-Lucent, which exhibit similar behaviors, with France Telecom used as the main illustration.Results for Alcatel-Lucent are provided in the appendix.
- Reference-price estimation: The reference price provides the center of the order book and determines the positions of the 2K limits.The construction must accommodate spreads that differ from one tick.
- Reference-price estimation: When the spread equals one tick, the reference price is the midprice; for wider spreads, the method uses the midprice when the spread is odd and an alternative nearby value when it is even.For even spreads, the selected value is the one closest to the previous reference price.
2.3 Model I: Collection of independent queues
Model I represents the limit order book as independent queues whose order-flow intensities depend on queue size, estimates those intensities from event data, and reproduces empirical queue distributions well.
- 2.3.1 Model I: Model I treats the book as 2K independent queues, with limit, cancellation, and market-order intensities determined by the target queue size.At each limit, the three order types are conditionally independent given the book state, and symmetric queues use identical intensity functions.
- 2.3.2 Empirical study: The estimation procedure records each queue event’s waiting time, event type, and pre-event queue size, then estimates intensities by maximum likelihood.Queue size is normalized using the stock’s average event size, and recording restarts whenever the reference price changes.
- 2.3.2 Empirical study: At Q±1, limit-order insertion is approximately constant but lower at zero, while cancellation increases concavely before flattening or slightly decreasing for larger queues.The lower insertion rate at zero is associated with the risk of creating a new best quote, whereas cancellation behavior is linked to queue priority.
- 2.3.3 Asymptotic behavior under Model I: Model I’s long-term queue behavior is completely determined by the arrival/departure ratio vector ρ, so different flow dynamics can share the same invariant distribution.The stationary distribution is derived explicitly for each queue.
- 2.3.3 Asymptotic behavior under Model I: The invariant distributions from Model I approximate the empirical queue distributions very well, supporting a mean-field explanation of order-book profiles.The comparison samples the book every 30 seconds and includes a Poisson-model benchmark.
2.4 Model II: Dependent case
Model II introduces dependencies into the queue-reactive order-book dynamics and shows that these extensions reproduce empirical queue distributions while linking order-flow intensities to neighboring queues and opposite-side liquidity.
- Model IIa: Empirical study: At Q±2, limit-order insertion decreases with queue size, while its behavior differs sharply according to whether Q±1 is empty.When q±1 = 0, insertion intensity rapidly reaches an asymptotic value; when q±1 > 0, it continues declining to a much lower value.
- Model IIa: Empirical study: At Q±2, cancellation is higher when Q±1 is empty, whereas market orders cannot arrive while Q±1 still contains limit orders.When Q±2 is the best limit, market-order intensity is similar to that at Q±1 but increases for queue sizes above 5 AES2.
- Model IIa: Asymptotic behavior: Theoretical joint distributions in Model IIa provide a very satisfying approximation to the empirical distributions of (q1, q2) for France Telecom.Model IIa is formulated as a quasi birth-and-death process, enabling matrix-geometric analysis of its asymptotic behavior.
- Model IIb: Empirical study: With an empty opposite queue, limit-order insertion is significantly larger, while abundant opposite-side liquidity is associated with more market orders.The paper relates these patterns to the relative attractiveness of prices and the temporary position of the efficient price.
- Model IIb: Asymptotic behavior: Model IIb’s joint distribution differs from empirical data when the reference price is fixed, but adding suitable reference-price moves makes the simulated density very close to the empirical one.The fixed-price model overproduces states where one first queue is empty, whereas empirical empty-queue states often trigger a reference-price change.
2.5 Example of application: Probability of execution
The paper uses its queueing models to estimate the probability that a trader’s limit order executes before the opposite best queue is depleted. The analysis incorporates queue priority and cancellation assumptions, while noting that these assumptions may overestimate execution probabilities.
- Execution probability is computed for a buy order submitted at Q−1, conditional on execution or total depletion of the opposite queue Q1.The order has size n0 and is initially behind orders already present at Q−1.
- Cancellation assumptions treat existing orders at Q−1 as equally likely to be canceled while protecting trader A’s order from cancellation.The cancellation intensity is also taken from the queue’s total cancellation intensity.
- The authors warn that execution probabilities may be slightly overestimated because lower-priority orders are actually more likely to be canceled.The required order-identifier data are unavailable for modeling this priority effect directly.
- For n0 = 1, the three queueing models give fairly similar execution probabilities, whereas a Poisson model with linearly increasing cancellation rates clearly overestimates them.
3 The queue-reactive model: a time consistent model with stochastic LOB and dynamic reference price
The queue-reactive model extends constant-reference-price queueing dynamics to a full period by allowing stochastic reference-price moves and optional order-book reinitialization. It is then used to analyze volatility, execution tactics, market impact, and scheduling slippage.
- 3.1 Building the model: Reference-price changes are triggered by spread insertions or depletion of a best queue, allowing the model to connect order-book dynamics with price movements.The model switches queues to neighboring values when the reference price changes.
- 3.1 Building the model: The reinitialization probability θreinit represents the share of price changes attributed to exogenous information and redraws the order-book state around the new reference price.
- 3.1 Building the model: With θreinit = 0, maximal mechanical volatility is 5 bps versus 14 bps empirically for France Telecom when θ = 1.The comparison suggests that order-book-driven dynamics alone generate less volatility than observed in this example.
- 3.2 Example of application: Order placement analysis: Theoretical scheduling slippage uses VWAP-based slice prices to assess scheduling quality while incorporating execution-induced market impact.
- 3.2 Example of application: Order placement analysis: T2 outperforms T1 under linear scheduling with a VWAP benchmark, whereas T1 slightly outperforms T2 under exponential scheduling with an arrival-price benchmark.Thus tactic performance depends on the associated scheduling strategy and benchmark.
- 3.2 Example of application: Order placement analysis: T2 obtains more passive execution but creates greater market impact because it remains in the queue longer, while T1’s impact is more instantaneous.The impact curves are concave in both time and volume.
4 Conclusion and perspectives
The paper models participant behavior through average responses to order-book states and turns those responses into a market simulator. The simulator supports pre-trade transaction-cost analysis for complex trading strategies, while historical order flow remains outside the framework.
- The model represents market-participant intelligence through average behaviors conditional on different limit-order-book states.
- The resulting market simulator supports investigation of transaction costs for complex trading strategies before execution.The authors describe this pre-trade cost analysis as simple and efficient within their framework.
- Historical order flow is not modeled, although its documented autocorrelation makes adding it a relevant direction for future work.
5 Appendix
The appendix supplies proofs, matrix formulations, numerical assumptions, confidence-interval procedures, and additional Alcatel-Lucent results for the queueing models. It also documents empirical quantities and figures used to assess the models.
- The appendix proves that the queueing process is V-uniformly ergodic, Harris positive recurrent, and convergent to an equilibrium distribution.
- Confidence intervals for Model IIb are approximated by neglecting possible intersections between the relevant queue-state events.
- Model IIa is formulated as a QBD process whose matrices encode upward, within-level, and downward transitions of the two-queue system.
- The appendix defines the stationary distribution π of the QBD process and states its normalization conditions.
- Additional figures report intensities, queue distributions, joint queue distributions, and Alcatel-Lucent results, while Table 5.5 reports estimated event-size measures.