Limit Order Book Dynamics Models
A limit order book is a stochastic queueing system with strategic participants. Prices change when liquidity at the best quote is depleted, but the route to depletion depends on limit-order arrivals, cancelations, market orders, hidden liquidity, and matching rules. A model that represents only OHLC bars cannot distinguish a quiet price move from a queue collapse. That distinction drives passive fill risk, short-horizon signals, and transaction costs.
The right model depends on its decision. A research feature may only need top-of-book imbalance; a passive simulator needs queue evolution; a market-making system needs joint dynamics of depth, flow, and inventory. Start from a deterministic event replay and documented feed semantics, as described in market microstructure.
State, events, and event time
Represent the book at event \(t\) by best prices, queue sizes, and optionally depth at multiple levels:
\[ Xt = (p^bt,p^at,Q^b{t,1:L},Q^a{t,1:L}, zt). \]
The event \(e{t+1}\) can add, cancel, or execute liquidity at a level. The state transition \(X{t+1}=T(Xt,e{t+1})\) is deterministic once the exchange event is known; the model estimates the conditional law of the next event. Event time avoids inventing idle observations during quiet periods.
| Model family | Core assumption | Strength | Main omission |
|---|---|---|---|
| Zero intelligence | Poisson arrivals with simple placement | Transparent baseline | State dependence |
| Queue-reactive | Intensities depend on queue state | Captures replenishment | Strategic heterogeneity |
| Markov state model | Discrete book state transitions | Fast for control | State discretization |
| Hawkes process | Events self- and cross-excite | Clustering | Queue mechanics unless added |
| Deep sequence model | Learns mapping from events | Flexible | Stability and interpretation |
A queue-reactive baseline
In a queue-reactive model, the intensity of each event type varies with current depth and imbalance. For example, a thin ask queue may receive more sell limit orders and more buy market orders, while a depleted bid may attract replenishment:
\[ \lambdak(Xt) = \exp(\thetak^\top \phi(Xt)). \]
Here \(\phi\) may include normalized depth, spread in ticks, recent signed flow, volatility, and intraday seasonality. Simulate the next event type from normalized intensities and its arrival time from an exponential waiting time. Price changes when a best queue reaches zero and the next occupied level becomes best.
import numpy as np
def sample_event(intensities, rng):
names = list(intensities)
rates = np.array([intensities[n] for n in names], dtype=float)
total = rates.sum()
wait = rng.exponential(1 / total)
kind = names[rng.choice(len(names), p=rates / total)]
return wait, kind
This code only samples timing and type. A credible simulator also samples size, price level, cancellation allocation, hidden executions, and session state. Never allow a negative queue or a crossed continuous book merely because the statistical generator produced one.
Features that connect state to price
The simplest top-of-book imbalance is
\[ \operatorname{OBI}=\frac{Q^b1-Q^a1}{Q^b1+Q^a1}. \]
It often predicts the next price move conditionally, but prediction weakens when queues are easily canceled. Order-flow imbalance (OFI) incorporates changes in quotes and displayed sizes, and can be more informative than static depth. Link both to order-flow imbalance signals, with features computed strictly from events available before the decision.
| Feature | Intuition | Execution use |
|---|---|---|
| Spread | Immediate crossing cost | Choose passive versus aggressive |
| Queue imbalance | Near-term pressure | Skew quote or urgency |
| Cancel-to-add ratio | Fragility of displayed depth | Discount displayed size |
| Trade intensity | Arrival opportunity/risk | Set participation cap |
| Microprice | Depth-weighted fair-price proxy | Price passive orders |
Calibration and validation
Calibrate intensities by instrument, venue, time of day, and regime. A pooled model usually learns that the open is different from midday but still misses event-driven states. Use rolling or expanding training windows, and reserve entire days—not random events—for testing. Random splitting leaks adjacent queue state into validation.
Validate distributions, not just one-step log likelihood. Simulated and observed data should agree on spread frequencies, queue-size distributions, inter-event times, price-change rates, autocorrelation of trade signs, and conditional fill probabilities. For a strategy simulator, the final validation is predicted versus realized fills and markouts by queue percentile.
Models with full-depth forecasts can look impressive while failing this test because the action is at the touch. Conversely, a simple two-queue model may be adequate if it calibrates passive outcomes and avoids false precision.
From model to trading decision
Use a book model to estimate distributions conditional on action. Posting a bid changes queue position and exposes the order to adverse selection; crossing the spread avoids that queue risk but pays immediacy. This is an execution-control problem, not a standalone forecast contest. Couple state forecasts with the queue and markout framework in inventory-risk market making.
Do not use future book updates to label a state, ignore sequence gaps, or assume trade prints consume visible depth. Those shortcuts create exceptionally good backtests and untradable live results.
Key takeaways
- Model the book in exchange event time, with queues and matching rules as first-class state.
- Queue-reactive intensities offer an interpretable middle ground between Poisson baselines and black boxes.
- Static imbalance is useful but cancelation, flow, and spread determine whether depth is durable.
- Validate simulated distributions and actual passive outcomes, not only predictive likelihood.
- The relevant forecast is conditional on the order you place and its resulting queue position.