Skip to content

Queue-Reactive (QR) Model

Algorithm description

The Queue-Reactive (QR) model (Huang, Lehalle, Rosenbaum 2015) represents the limit order book as a continuous-time Markov jump process whose event intensities depend on the current observable state of the book. Three event types modify the book at each price level:

  • Limit order insertions (Add)
  • Cancellations (Cancel)
  • Market orders (Trade)

For each event type \(e\) on queue \(q_i\), the arrival intensity is a function of the book state:

\[\lambda^e = \lambda^e(q_{-K}, \ldots, q_{-1}, q_1, \ldots, q_K)\]

In the simplest version each queue is independent — \(\lambda^e(q_i)\) — which makes estimation trivial but ignores bid–ask coupling. The original paper extends this to \(\lambda^e(q_1, q_{-1})\), conditioning on both best queues, at the cost of aggressive binning to get enough observations per bin.

The price reference \(p_\text{ref}\) is shifted by a separate stochastic mechanism when the best queue is depleted or a new level is created inside the spread.

Why it matters

QR is a workhorse for interactive LOB simulation used by market makers, prop trading firms, and brokerages. Its key virtues over zero-intelligence models (Smith et al. 2003; Cont–De Larrard 2013):

  • Data-driven: transition rates estimated empirically, no ad-hoc parameter tuning.
  • Interpretable: each rate is the empirical event frequency under an observable state.
  • Captures queue dependence: event intensities respond to book asymmetries.

Known limitations (motivating extensions)

  1. Exponential inter-event times: a continuous-time Markov jump process generates exponential waiting times by construction. Real markets exhibit pronounced clustering at exchange round-trip latency (latency races account for ~20% of volume per Aquilina et al. 2021) — QR averages this out entirely.
  2. No post-execution market impact: after a metaorder completes, the book in QR evolves as if nothing happened. For strategy evaluation this is a critical flaw — simulated P&L systematically overstates profit and understates risk.
  3. Sparse state space under heavy conditioning: conditioning on full queue-size tuples leaves most cells poorly populated, forcing coarse binning.

Extensions

The most practical recent extension is reality-gap-lob-simulation (Noble, Rosenbaum, Souilmi 2026), which:

  • Projects the state onto \((\text{Imb}, n)\) — volume imbalance + spread — making estimation tractable with a single scalar of bid-ask coupling.
  • Replaces the exponential inter-event distribution with a flexible empirical one, capturing the latency-race mode.
  • Adds a power-law market-impact feedback kernel so post-execution book dynamics reflect the metaorder's footprint.
  • Uses random order volumes drawn from state-conditional empirical distributions rather than unit sizes, so queue depletion arises naturally.

When to use / when not to use

Use when: - You need an interactive simulator for backtesting or strategy stress tests. - The asset is large-tick and the book state is well-summarised by imbalance + spread. - You want interpretable, data-driven transition rates.

Avoid or adapt when: - Modelling small-tick assets where \(n \gg 1\) is common — the imbalance projection breaks down. - Studying mechanisms that require non-Markov memory (e.g., persistent meta-order impact) — extend with explicit feedback kernels. - Modelling latency races explicitly — you need non-exponential inter-event times.

Implementations

  • Huang–Lehalle–Rosenbaum 2015 code (original Markov QR).
  • Noble–Rosenbaum–Souilmi 2026 simulator — extensions described above; code on GitHub.

Connections