Limit Order Book (LOB)

• The limit order book (LOB) is a continuous double-auction mechanism that records active buy and sell orders, their prices, quantities, and submission times, determining displayed and latent liquidity in global financial markets for stocks, future, forex etc.
• The accuracy of LOBs in describing true liquidity levels is limited by the inclusion of orders that are hidden, cancelled, or posted in dark pools, making empirical modeling challenging for certain market settings
• LOB data show clustering across time, heavy-tailed volume of limit orders, and order-dependent market execution needing further investigation to solve the enigmatic nature of financial markets

A limit order book (LOB) is a continuous double-auction mechanism that records active buy and sell orders, their prices, quantities, and submission times. Traders may demand immediacy by submitting orders that execute against existing liquidity or supply liquidity by posting orders that wait in the book. The best bid and best ask determine the bid–ask spread and mid-price, while price–time queues, executions, cancellations, hidden liquidity, and order-flow dynamics jointly determine displayed liquidity and short-term price formation. LOB mechanisms are used in more than half of global financial markets (Gould et al., 2010).

1. Structure and terminology

An order can be represented as

$x=(p_x,\omega_x,t_x),$

where $p_x$ is the order price, $\omega_x$ its signed size, and $t_x$ its submission time. Under the convention used in the LOB literature, $\omega_x>0$ denotes a sell order and $\omega_x<0$ a buy order. The absolute size $|\omega_x|$ is the quantity offered or requested.

A sell order commits its owner to sell up to $|\omega_x|$ units at a price no lower than $p_x$. A buy order commits its owner to buy up to $|\omega_x|$ units at a price no higher than $p_x$0. The lot size $p_x$1 is the smallest permitted trade quantity, while the tick size $p_x$2 is the smallest permitted price increment:

$p_x$3

The pair $p_x$4 constitutes the market’s resolution parameters. These parameters affect queue formation, price competition, order splitting, spread size, and the frequency of price changes.

The active book at time $p_x$5 is

$p_x$6

Its buy and sell subsets are

$p_x$7

At each price, orders form a queue. Under price–time priority, better prices have priority, and orders at the same price are ranked from oldest to newest. Thus buy orders are ranked from highest to lowest price and then by submission time; sell orders are ranked from lowest to highest price and then by submission time.

An incoming order that immediately matches existing liquidity is conventionally called a market order. An order that does not immediately match and instead rests in the book is called a limit order. These labels describe execution status rather than fundamentally different objects: a priced order may be aggressive enough to execute immediately or passive enough to wait.

The LOB is not a complete representation of desired trades. Traders may withhold intentions, use hidden or iceberg orders, trade in dark pools, or submit orders only when immediate execution is desired. Consequently, displayed depth is only a partial measure of latent supply and demand (Gould et al., 2010).

2. Quotes, spread, depth, and liquidity

The best bid is the highest active buy price,

$p_x$8

and the best ask is the lowest active sell price,

$p_x$9

The bid–ask spread and mid-price are respectively

$\omega_x$0

Under ordinary continuous trading, $\omega_x$1, so the spread is nonnegative. Special mechanisms, including bilateral trading restrictions or auctions, can temporarily violate this property.

The spread measures the market’s valuation of immediacy and execution certainty. A market order executes quickly but pays the spread and may walk through several price levels. A limit order may obtain a better price but faces uncertain execution, queueing delay, and cancellation risk.

Because absolute prices change over time, empirical research often uses relative prices. For a price $\omega_x$2,

$\omega_x$3

is its distance below the bid, while

$\omega_x$4

is its distance above the ask. For an order $\omega_x$5,

$\omega_x$6

Negative relative prices lie inside the spread and improve the corresponding best quote. Nonnegative relative prices are at or behind the best quote on their side.

Bid-side depth at price $\omega_x$7 is

$\omega_x$8

and ask-side depth is

$\omega_x$9

Since buy sizes are negative under the signed convention,

$t_x$0

The absolute amount of buy liquidity at $t_x$1 is $t_x$2, while sell liquidity is $t_x$3. A depth profile is the collection of price–depth pairs. Researchers generally use relative depth because bid and ask prices move:

$t_x$4

Liquidity has three broad dimensions:

1. Tightness: the cost of trading immediately, closely related to the spread.
2. Depth: the amount of order flow required to move prices by a given amount.
3. Resiliency: the speed with which the book and prices recover after a shock.

These dimensions are distinct. A narrow spread does not guarantee substantial depth, and a deep book may recover slowly after a large market order (Gould et al., 2010).

3. Matching, execution, and price formation

When an order arrives, the matching algorithm searches for eligible opposite-side orders. A buy order is executable against sells when

$t_x$5

and a sell order is executable against buys when

$t_x$6

The incoming order is matched to the highest-priority eligible opposite-side order. If its quantity exceeds the first resting order, the remainder continues through the queue and potentially through successive price levels. Any unfilled residual becomes active at its stated price, provided it is no longer executable. Trades generally occur at the price of the resting order rather than necessarily at the incoming order’s submitted price.

A sufficiently large market order “walks the book.” If a sell market order consumes all buy quantity from the best bid downward through several prices, the new bid becomes the highest remaining eligible buy price. The analogous mechanism applies to a buy market order consuming ask liquidity.

Cancellations are central to LOB dynamics. An active order remains in the book until it executes or is cancelled, and empirical estimates commonly find that most active orders are cancelled rather than executed. Cancellations depend on order distance, queue position, the surrounding book, and recent order flow. They may increase after same-side market orders, when traders probe for hidden liquidity, or after new higher-priority orders make existing quotes less attractive.

Hidden liquidity includes iceberg orders, entirely hidden limit orders, dark pools, and undisplayed portions of otherwise visible liquidity. Displayed depth therefore does not measure actual executable supply and demand exhaustively.

Price changes arise mechanically from order arrivals, executions, and cancellations. For an incoming buy order:

• if $t_x$7, it rests behind existing buys;
• if $t_x$8, it rests inside the spread and raises the bid;
• if $t_x$9, it executes against sells and may raise the ask.

The sell-side cases are symmetric. Price analysis depends strongly on sampling convention. Observations may be indexed by clock time, event time, or trade time, and a pattern visible at one sampling frequency may disappear or reverse at another.

Opening and closing auctions use a different mechanism. Orders accumulate without execution, and the auction price is selected to maximize executable volume:

$\omega_x>0$0

where $\omega_x>0$1 is the volume executable at price $\omega_x>0$2. Auction trades occur at the common uncrossing price $\omega_x>0$3.

A mathematical framework developed by Baldacci, Bergault, and colleagues separates stochastic order flow from deterministic market clearing. In that formulation, order flow first creates a potentially crossed transient configuration, after which a clearing operator matches compatible buy and sell mass and returns the state to the admissible LOB space. The generator factorizes as

$\omega_x>0$4

where $\omega_x>0$5 is the generator of unconstrained order flow and $\omega_x>0$6 is the clearing operator (Cont et al., 2023).

4. Empirical regularities

LOB data exhibit recurring but market-dependent statistical regularities. Order sizes are highly heterogeneous, cluster at round numbers such as $\omega_x>0$7, $\omega_x>0$8, and $\omega_x>0$9, and may have heavy-tailed aggregate distributions. Very large market orders rarely walk through many price levels. Relative order prices often display power-law-like tails, with reported exponents differing across markets.

Mean relative depth profiles are commonly hump-shaped: depth rises over the first few price levels, reaches a maximum away from the best quote, and then declines deeper in the book. The hump reflects a trade-off between execution probability and price improvement. Orders near the best quote have a higher chance of execution but poorer prices; orders farther away offer better prices but have lower execution probability.

Events cluster in time. Buy market orders, inside-spread limit orders, and cancellations tend to occur more frequently after events of the same class than unconditional frequencies would imply. Possible explanations include strategic order splitting, imitation, common information, undercutting, queue competition, and repeated algorithmic probing.

Order flow is state-dependent. Arrival behavior depends on the current spread, best-quote depth, deeper liquidity, recent order flow, volatility, time of day, recent inter-arrival times, and the side and type of the most recent event. The precise relationships are not universal: high volatility may increase market-order activity in some datasets and limit-order activity in others.

Mid-price returns are generally negatively autocorrelated at very short horizons and approximately uncorrelated at longer horizons. The duration of short-run negative autocorrelation has shortened in faster markets. Returns have heavier tails than a normal distribution over horizons from seconds to days. At short horizons, tails are often approximated by a power law with exponent near $\omega_x<0$0:

$\omega_x<0$1

At longer horizons, returns become increasingly close to Gaussian, a phenomenon known as aggregational Gaussianity.

Absolute or squared returns display long memory, often represented as

$\omega_x<0$2

Equivalently, if

$\omega_x<0$3

then ordinary short-memory behavior has $\omega_x<0$4, whereas long memory has $\omega_x<0$5. Reported volatility Hurst exponents range from roughly $\omega_x<0$6 to $\omega_x<0$7, depending on market and estimator.

Order flow itself may also be long-memory: buy and sell signs, limit-order prices, best-quote depth, market-order arrivals, and cancellations have all been reported to show persistence. Returns generally do not inherit this long memory because predictable order flow is absorbed by changing liquidity. When one-sided market orders are likely, the book may become deeper on the vulnerable side, reducing their price effect (Gould et al., 2010).

Market impact is generally concave in market-order size. A common empirical specification is

$\omega_x<0$8

where $\omega_x<0$9 is mean logarithmic mid-price impact. Reported exponents range roughly from $|\omega_x|$0 to $|\omega_x|$1, although logarithmic impact can fit better in some datasets and power laws may overestimate the effect of very large orders.

Trade imbalance is related to price changes, but order-flow imbalance—which also includes limit-order arrivals and cancellations—typically produces stronger relationships. Price sensitivity is higher for stocks with smaller best-quote depth, consistent with thinner books being more vulnerable to order flow.

5. Models and scaling limits

LOB models occupy a spectrum from zero-intelligence stochastic models to strategic, agent-based, and infinite-dimensional representations.

Perfect-rationality and game-theoretic models treat traders as optimizing utility, profit, execution, or information objectives. They illuminate fill probabilities, informed trading, cancellations, optimal execution, and strategic placement. Their limitations include unobservable valuations and beliefs, strong equilibrium assumptions, restricted price grids, and difficulty reproducing continuous, nonstationary order flow.

Zero-intelligence models specify stochastic arrivals, cancellations, and price locations without explicitly optimizing trader objectives. Poisson models are tractable and can produce expressions for mean spread, expected depth, price diffusion, and queue distributions. However, independent arrivals contradict clustered order flow, uniform price placement contradicts empirical relative-price distributions, and infinitesimal tick limits eliminate genuine price-level queues.

Hawkes-process extensions allow past events to excite future events. A generic intensity is

$|\omega_x|$2

These models reproduce clustered high- and low-activity periods and mutual excitation among event types, although they generally impose event-type structure exogenously.

Queue-based diffusion limits provide analytical approximations for best bid and ask queues. Cont and de Larrard model the queue state in the positive orthant, with order-flow increments accumulating until one queue reaches zero. The interior limit is a correlated Brownian diffusion, while depletion causes a jump into the interior through a queue reinitialization distribution. The limiting process yields tractable probabilities of the next price move and distributions of the time until the next price change (Cont et al., 2012).

A fluid law-of-large-numbers limit produces a deterministic coupled ODE–PDE system for best prices and full buy and sell volume densities. Under appropriate scaling, tick size, event time, individual volume impact, and inter-event time vanish while arrival rates increase. The stochastic LOB converges in probability to deterministic best-price trajectories and transport-reaction equations for liquidity profiles. This is a fluid limit rather than a diffusion approximation (Horst et al., 2015).

SPDE models represent the book as a signed density centered at the mid-price. A representative equation on the ask side is

$|\omega_x|$3

The terms represent price-space diffusion, convection, proportional cancellation, new order submissions, and multiplicative order-flow noise. Under invariant-subspace conditions, the infinite-dimensional system reduces to a finite-dimensional Markov diffusion. Principal-eigenfunction models produce parsimonious factors for bid and ask volume, depth, imbalance, and price (Cont et al., 2019).

Conservation-law models instead describe order migration between price levels. If $|\omega_x|$4 denotes resting quantity and $|\omega_x|$5 its flux, conservation is expressed as

$|\omega_x|$6

where $|\omega_x|$7 represents interior sources and sinks. Nonlinear migration produces self-similar recovery after aggressive liquidity taking, with touch displacement and recovery width scaling as

$|\omega_x|$8

This mechanism differs from models that restore each price level independently (Rosenzweig, 2019).

6. Market structure, execution, and applications

The LOB is simultaneously a storage system for future supply and demand, a price-discovery mechanism, a venue for strategic information revelation, and a source of endogenous liquidity and fragility.

Queue position is particularly important under price–time priority. The execution prospects of two orders at the same price can differ substantially according to the quantity ahead of each order. A queue-position model represents best bid and ask queues together with the position $|\omega_x|$9 of a particular order. Under uniform cancellation,

$|\omega_x|$0

so market orders remove volume from the front of the queue and cancellations reduce position proportionally. Fluid and diffusion limits yield deterministic execution-time approximations and Gaussian fluctuations around those approximations (Guo et al., 2015).

Execution models balance spread cost, timing risk, adverse selection, and market impact. Large orders may be split across time, but order splitting exposes traders to changing liquidity, information leakage, and adverse price movements. Since order-flow imbalance includes passive orders and cancellations, execution models that use trades alone omit relevant liquidity dynamics.

Market-making models must preserve consistency between orders and prices. A weakly consistent LOB model restricts price movements to directions permitted by the arriving order, imposes price changes only at order-event times, and treats liquidity provision through switching and inventory impulses. Its market-making problem is formulated as a Hamilton–Jacobi–Bellman quasi-variational inequality with a jump-process generator and intervention operator. The model emphasizes that independent exogenous prices can create economically impossible events, such as a buy market order coinciding with a downward ask movement, and can consequently overstate market-making profitability (Law et al., 2019).

Empirical execution studies show that the previous event can affect subsequent submission probabilities even after conditioning on current imbalance. A discrete Markov model with queue imbalance and the most recent order type can therefore outperform a model based only on current queue volumes. In one numerical framework, the action set includes waiting, submitting a limit order, submitting a market order, and cancelling a limit order. The resulting policy balances non-execution risk against adverse-selection risk (Gonzalez et al., 2017).

LOBs also differ across market structures. In a quasi-centralized LOB, bilateral counterparty-credit limits mean that each institution sees and can trade against only a filtered local book. The global book may contain buy and sell orders that cross without trading, and the global spread can therefore be negative. Trade-relative coordinates may provide more stable statistical regularities than quote-relative coordinates because recent trade prices are common informational benchmarks whereas executable quotes are institution-specific (Gould et al., 2015).

Outside financial markets, LOB mechanisms have been proposed for transactive-energy systems. Such designs are intended to support price discovery involving wholesale electricity prices, distribution-system constraints, distributed-energy-resource constraints, and reservation prices. The supplied description remains high-level and does not specify a complete clearing algorithm, settlement rule, network-feasibility formulation, or delivery-guarantee mechanism (Sreekumar et al., 2023).

7. Measurement, simulation, and unresolved issues

LOB analysis is difficult because observations are irregularly timed, markets are nonstationary, order flow is dependent, latency and data-refresh rules obscure the state observed by traders, and hidden or off-book liquidity is missing. Sampling frequency changes measured volatility and autocorrelation, while exchanges differ in tick size, lot size, matching rules, and priority systems.

Power-law claims require particular caution. Log–log ordinary least squares can produce biased exponents, continuous-data estimators are inappropriate for discrete quantities, and lower cutoffs must be estimated rather than assumed. Long-memory estimates are also fragile because finite samples and regime changes can mimic asymptotic dependence.

Simulation models are used for backtesting, stress testing, strategy training, and execution-policy design. Their validity depends on reproducing both unconditional stylized facts and the response of the book to an external trader’s orders. A simulator that omits market impact may generate strategies that perform well only because their own orders do not alter prices, liquidity, or subsequent order flow. Synthetic data should therefore be combined with genuine out-of-sample market data whenever possible (Jain et al., 2024).

Recent reconstruction models infer deeper book volumes from trades-and-quotes data. The Limit Order Book Recreation Model predicts volumes at levels $|\omega_x|$1 through $|\omega_x|$2 using rolling TAQ histories, sparse positional encoding, recurrent modules, and time-aware temporal dynamics. Its chronological evaluation avoids the lookahead bias associated with random train–test splits, although it reconstructs a sampled partial book rather than the complete event-by-event LOB (Shi et al., 2021).

Several unresolved questions remain:

1. Which empirical regularities arise from matching rules, strategic behavior, heterogeneous beliefs, order splitting, hidden liquidity, or external information?
2. How do stylized facts change in modern high-speed markets and across venues?
3. How should equilibrium concepts be adapted to nonstationary books subject to continual shocks?
4. How can volatility and liquidity measures incorporate the full state of the book rather than only bid, ask, or mid-price histories?
5. What determines liquidity resiliency after market-order shocks?
6. Under what conditions does algorithmic trading improve liquidity or amplify instability?
7. How should liquidity be measured across fragmented venues and credit-constrained local books?
8. How can instantaneous and permanent market impact be identified causally?
9. Can a single model reproduce event-time microstructure, heavy-tailed returns, volatility clustering, long-memory order flow, queue dynamics, hidden liquidity, and cross-market liquidity?

No existing model reproduces all major empirical facts simultaneously. Tractable models often rely on assumptions contradicted by data, while highly strategic models may fit selected patterns using unobservable information and utility structures. The principal methodological direction is therefore integration: exact exchange mechanics, realistic order-flow dependence, cancellations and hidden liquidity, heterogeneous and adaptive traders, nonstationarity, and statistically careful validation across multiple time scales and trading venues.