6.1.3 The Limit Order Book

Basic Prerequisites: 02 · Limit vs Market Orders.

---

1. Intuition & Practical Objective

If page 02 was about the actions, this page is about the board. The limit order book is a state: a complete, ordered description of every resting commitment. Everything a pricing or execution engine computes - spread, depth, microprice, imbalance, queue position - is a function of this state. Get the state representation right and the rest is arithmetic; get it wrong and every downstream number is quietly corrupted.

The practical objective is to answer three questions precisely:

1. What exactly is in the book? (the objects and the L1/L2/L3 data levels)
2. How is it summarised? (the state vector and its derived quantities)
3. How does it evolve? (the book as a stochastic process, and what empirical shape it takes)

The one-sentence essence. “The book is a high-dimensional queueing state - a per-price queue of unexecuted orders - and market data feeds are lossy projections of it: L1 keeps the touch, L2 keeps price-level totals, L3 keeps every order. Only L3 retains queue position.”

---

2. Mathematical Ground Truth & Derivations

2.1 The state, exactly. Index resting orders by $i$. In full generality (L3),

$${\mathcal{B}}_t=\{i:\text{order }i\text{ is resting at time }t\},i\mapsto ({\text{side}}_i,p_i,q_i,t_i^{\text{arr}},{\text{id}}_i).$$

Aggregating by price gives the depth functions - the probabilistic content that most models use:

$$D_t^b(p)=\sum _{i:\text{buy},p_i=p}{q}_i,D_t^a(p)=\sum _{i:\text{sell},p_i=p}{q}_i,p\in \delta \mathbb{Z},$$

on a tick grid $\delta$ (e.g.\ $\delta =$$0.01). The best levels and the standard summaries are then pure functionals:

$$b_t=\max \{p:D_t^b(p)>0\},a_t=\min \{p:D_t^a(p)>0\},s_t=a_t-b_t,m_t=\frac{1}{2}(a_t+b_t).$$

2.2 The data levels are nested projections.

Level	Content	Retains	Loses
L1	best bid/ask + sizes $(b_t,q_t^b,a_t,q_t^a)$	the touch	everything away from the touch
L2	aggregated size at the top $N$ prices per side	depth profile	who is in the queue
L3	every order event (add/modify/cancel/execute) with ids	queue position, per-order identity	-

The nesting matters because queue position is not a function of L2. Two books with identical L2 pictures can have completely different L3 structures - one order of 500 lots, or 500 orders of 1 lot - and a resting order faces different fill odds in each. Queue position is the single most valuable piece of information that only L3 preserves.

2.3 Derived signals. Beyond the spread and mid, the tradeable summaries are:

• Book imbalance (a size-weighted touch signal): $I_t=\frac{q_t^b-q_t^a}{q_t^b+q_t^a}\in [-1,1]$.
• Microprice (the size-weighted touch, which leans away from the heavier side because that side will likely be consumed):

$$m_t^{\text{micro}}=\frac{q_t^b{a}_t+q_t^a{b}_t}{q_t^b+q_t^a}.$$

If the bid is large ($q_t^b\uparrow$) the microprice rises toward $a_t$: heavy resting demand signals upward pressure. Empirically $m^{\text{micro}}$ forecasts the next mid change better than $m_t$ - it is the simplest “micro-price”.

• Depth profile / book shape: the cumulative function $p\mapsto {\sum }_{p^{\prime }\le p}{D}_t^b(p^{\prime })$; its curvature is an empirical object (§2.4).

2.4 The book as a stochastic process. Model the grid as a state vector ${\mathbf{x}}_t=(D_t(p_k){)}_k$; the book evolves by a superposition of events, each a random state transition:

$$\mathbf{x}\stackrel{\text{limit add at }(p,\eta )}{\to }\mathbf{x}+\eta e_p,\mathbf{x}\stackrel{\text{cancel/execute }(p,\eta )}{\to }\mathbf{x}-\eta e_p,\mathbf{x}\stackrel{\text{market order}}{\to }\text{deplete best levels}.$$

Cont, Stoikov & Talreja (2010) formalise this: the book is a continuous-time Markov chain in the queue sizes, with arrival intensities that depend on the price’s distance from the touch. That distance dependence is the empirical key (Bouchaud, Mézard & Potters 2002): limit-deposit and cancellation rates are roughly independent of price-level index, so cumulative depth grows approximately linearly in the number of levels - a concave, roughly linear-in-distance book shape. Far from the touch, the book is a nearly homogeneous “reservoir”; at the touch it is the thin, contested layer that sets the price.

---

3. Computational Implementation - L3 retained, L2 derived, invariants checked

The core lesson of §2.2 in code: keep every order (L3), derive the L2 view as a pure function, and assert the invariants that keep the state sane. Standard library only.

The L3 map shows what L2 hides: at $100.00 there are two orders (#1 = 200 lots, #2 = 100 lots) and their order matters for fill priority. L2 shows only “300”. The two load-bearing invariants are (i) the book is never crossed ($b_t<a_t$ at rest - a crossing triggers a match, so it cannot persist), and (ii) L2 totals must equal the L3 sums at every level (a mismatch means the reconstruction has drifted).

---

4. Failure Modes & First-Principles Breakdowns

1. L2 is not enough to compute queue position. Because queue position is an L3 object (§2.2), any strategy that reasons about fills from an L2 feed is reasoning about the aggregate, not the queue it actually holds. This is the single most common production mistake, and the reason exchanges sell L3 feeds.
2. The book you hold is a reconstruction, not the truth. The venue’s book is authoritative; yours is rebuilt from a message stream. A dropped, duplicated, or out-of-order event yields a state that looks plausible (spread positive, sizes sane) but is wrong - silently mispricing everything derived from it. Recovery requires periodic snapshots and sequence-number gap detection (LOB Mechanics & L3 Data).
3. Snapshot staleness. Even a correct book is correct as of its last event. Between the exchange’s clock and yours, the state has moved - a hazard that becomes a loss mechanism under latency (§05).
4. Uncrossed ≠ unmanipulated. The invariant $b_t<a_t$ holds mechanically, but a book can be deliberately shaped - spoofing layers, fleeting liquidity - so a “valid” state is not a truthful one. The state machine cannot distinguish genuine intent from committed-then-cancelled orders.
5. Aggregate depth misreads fragility. Two identical L2 snapshots can hide one large order or a thousand small ones. The first evaporates when its owner cancels; the second needs a thousand independent decisions. Resilience (FPR Ch 1) depends on this microstructure that L2 erases.

---

5. References

• Hasbrouck, Empirical Market Microstructure
• Foucault, Pagano & Röell, Market Liquidity
• Bouchaud, Mézard & Potters (2002), Statistical properties of stock order books, Quantitative Finance 2(4), 251–256
• Cont, Stoikov & Talreja (2010), A stochastic model for order book dynamics, Operations Research 58(3)
• Gould, Porter, Williams, McDonald, Fenn & Howison (2013), Limit order books, Quantitative Finance 13(11)

---

6. Connected Graph Bridges

• Back: 02 · Limit vs Market Orders · Index Hub
• Forward: 04 · Matching & Priority · 06 · Advanced Extensions
• Related: Limit Order Book Mechanics & L3 Data · Market Microstructure & Order Types · Ergodicity & Statistical Mechanics