Trade Arrival Dynamics and Quote Imbalance in a Limit Order Book¶

Authors: Alexander Lipton, Umberto Pesavento, Michael G. Sotiropoulos Institution: Bank of America Merrill Lynch; Lipton also Imperial College Year: 2013 (December 2013) arXiv: 1312.0514 Categories: q-fin.TR

Plain-language abstract¶

Every trader knows that when the bid queue is much bigger than the ask queue, the next price move is more likely to be up. This paper turns that intuition into a semi-analytic closed-form probability for the next mid-price move (or trade arrival) conditional on the current top-of-book imbalance. The machinery: model queues as correlated Brownian motions whose first crossing of zero defines a price move; extend to a third dimension for near-side trade arrival; solve the resulting hitting-probability PDE via polar / spherical coordinate transforms. Calibrated on Vodafone (VOD.L) Q1 2012 data. A practical side-finding: passive orders lose ~60% of their theoretical spread capture purely from queue-depletion / trade-flow interaction — before any adverse-selection considerations.

Key contributions¶

Formal definition of quote (book) imbalance as a tradable signal:

$$I = \frac{q^b - q^a}{q^b + q^a} \in [-1, 1]$$

Positive ⇒ bid side heavier ⇒ next mid-move biased upward.

Empirical stylised facts (VOD.L, Q1 2012):
Average next-mid move is a nearly linear function of $I$, up to ~⅓ of the spread at high imbalance.
Average waiting time to the next mid move is decreasing in $|I|$ — imbalanced books move faster.
The average price move before the next trade differs by side: a buy trade sees on average a lower mid-price than a sell trade from the same starting imbalance. This is the signature of queue-depletion-driven slippage (plus adverse selection).
Two-dimensional diffusion model for queues:

$$dq^b = \mu_b \, dt + \sigma_b \, dw^b, \qquad dq^a = \mu_a \, dt + \sigma_a \, dw^a$$

with $\text{corr}(dw^b, dw^a) = \rho_{xy}$ (typically negative). Either queue hitting zero triggers a tick in the corresponding direction; a fresh queue is resampled from an empirical replenishment distribution.

Semi-analytic up-tick probability. Solves the hitting-probability PDE via two coordinate transforms:
decorrelate $(x, y) \to $ principal axes;
convert to polar $(r, \phi)$.

In polar coordinates the PDE reduces to $P_{\phi\phi} = 0$ with boundary conditions $P(0) = 0$, $P(\varpi) = 1$ where $\cos \varpi = -\rho_{xy}$. Solution:

$$P(\phi) = \frac{\phi}{\varpi}$$

In the original queue coordinates this yields $P_{\uparrow}(q^b, q^a) = \phi(q^b, q^a)/\varpi$ — a closed-form conditional up-tick probability depending only on current queues and one fitted correlation.

Three-dimensional extension for near-side trade arrival. Adds an unobserved diffusion $\phi$ whose zero-crossings mark trade arrivals, correlated with both queue processes. The hitting-probability PDE is solved semi-analytically via a sphere-to-strip coordinate change and a generalized Fourier series in $(\phi, \theta)$-space. Expansion coefficients obtained by matrix inversion $c = J^{-1} I$.
Passive-fill slippage finding. Calibration shows that ~60% of the theoretical spread capture of a passive fill is lost purely from queue-depletion / trade-flow interaction — even before invoking adverse selection. Provides a pure-microstructure explanation for why posting passively is less profitable than naïve analysis suggests.
Practical decision rule for brokers: at moderate imbalance, keep the order resting; at extreme imbalance (probability of unfavourable move ~90%), cross the spread. Optimal spread-crossing policies deferred as future work.

Method summary¶

Signal definition¶

\[I = \frac{q^b - q^a}{q^b + q^a}\]

where $q^b, q^a$ are the best-bid and best-ask posted quantities.

2D queue model & hitting probability¶

With Brownian queue dynamics and first-zero-crossing triggering a tick, the up-tick probability satisfies the time-independent diffusion PDE with boundary $P = 1$ on the $q^a = 0$ axis, $P = 0$ on the $q^b = 0$ axis. After decorrelation and polar transform, the solution is $P(\phi) = \phi / \varpi$ where:

Transform step	Effect
Rotate axes to principal components of $\rho_{xy}$	removes queue correlation
Polar $(r, \phi)$	reduces PDE to $P_{\phi\phi} = 0$
Closed-form $P(\phi) = \phi / \varpi$	plug back into original $(q^b, q^a)$

3D model: add trade arrival¶

Adds a third SDE $d\phi = \mu_\phi \, dt + \sigma_\phi \, dw^\phi$ whose zero-crossings mark near-side trade arrivals. Correlations with queue processes ($\rho_{xz}$, $\rho_{yz}$) encode how aggressive flow and queue depletion are linked.

PDE becomes 3D; solved via spherical coords, strip-coord transform, then generalized Fourier series. Boundary conditions select the event of interest: - favourable price move (for a broker); - unfavourable price move; - near-side matching trade.

Calibration¶

Asset: VOD.L (Vodafone, LSE).
Period: all trading days Q1 2012.
Parameters fitted: initial $\phi_0$, correlations $\rho_{xy}, \rho_{xz}, \rho_{yz}$ — four scalars.
Fit targets: joint fit of empirical next-move mid-price bias, empirical trade waiting times, and empirical event probabilities, each bucketed by imbalance.

Main results¶

Closed-form up-tick probability $P_{\uparrow}(q^b, q^a) = \phi / \varpi$ in the 2D model matches empirics well at moderate imbalance.
3D model additionally reproduces the buy/sell asymmetry in waiting-time and mid-price-bias curves — a property no 2D queue-only model can capture.
Model qualitatively captures the steep decrease in trade arrival time at extreme imbalance, though quantitatively underestimates it in the tails.
Unfavourable-move probability exceeds 90% at high imbalance from the perspective of a passive-near-side broker — strong signal to cross the spread in that regime.
Passive fills capture only ~40% of the quoted spread on average — the other 60% dissipates via queue-depletion/trade-flow coupling that the 3D model mechanistically explains.

Limitations¶

Single stock (VOD.L), one quarter. Cross-sectional robustness not tested in this paper (follow-up work by the authors and others covers more).
Diffusive queues — real queue arrivals are point-process-like (Hawkes-style clustering). The diffusion limit is an approximation that washes out sub-tick clustering.
Unobserved trade-arrival process $\phi$ is a modelling device; the calibration gives a fit but the process is not directly identified from data.
Steep-imbalance tails under-predicted: trade arrivals come much faster at $|I| \to 1$ than the diffusive model anticipates.
No optimal spread-crossing policy — deferred to future work; the paper provides the probabilities a policy would be built on.
Next-level dynamics not modelled — when a queue depletes, replenishment is sampled from an empirical distribution rather than a structural deeper-level model.

Connections¶

Quote Imbalance $I$ is the second OBI-family signal in this wiki alongside the event-based OFI (Cont–Kukanov–Stoikov). Key distinction: OFI cumulates signed event contributions over an interval; QI is the instantaneous snapshot ratio of best-bid vs best-ask quantities. Both are predictive short-horizon signals; QI is much simpler to compute in real time.
Microprice connection: QI enters Stoikov's microprice definition as the weight of the spread-adjusted mid.
Queue-Reactive Model: queue-reactive-model uses QI-like state projection but with discrete events rather than a diffusive limit. Noble–Rosenbaum–Souilmi 2026 (reality-gap-lob-simulation) explicitly build on the $(I, n)$ state — this paper's queue-imbalance work is a direct antecedent.
Passive-fill slippage: the ~60% lost-capture finding is a foundational quantitative argument for market-maker inventory modelling. Connects to market-making and optimal-execution.
Companion / contrast with price-impact-order-book-events: CKS give the OFI → mid-price linear impact coefficient; this paper gives the event probability given current QI. The two pieces combine into a full execution model (expected value vs variance of next-tick cost).
Classical reference: Cao, Hansch & Wang (2009) introduced depth imbalance (QR) and width imbalance (HR) as related signals — not currently in the wiki but referenced across the OFI literature.

Transform step	Effect
Rotate axes to principal components of \(\rho_{xy}\)	removes queue correlation
Polar \((r, \phi)\)	reduces PDE to \(P_{\phi\phi} = 0\)
Closed-form \(P(\phi) = \phi / \varpi\)	plug back into original \((q^b, q^a)\)