Agent-Based LOB Simulator Calibration for Slippage Stress Testing

Date: 2026-02-28
Category: research (slippage modeling)

Why this matters

Backtests usually assume a static impact curve and average fill behavior. Real execution risk, however, is generated by interacting agents: market makers widen spreads, toxic flow accelerates cancellations, and latency pockets create bursty adverse selection.

An agent-based limit-order-book (LOB) simulator gives you a way to answer:

• “How bad can slippage get when liquidity providers de-risk together?”
• “Which execution tactic survives cancel storms?”
• “Do my controls still work when microstructure feedback loops become nonlinear?”

This playbook shows how to build a calibrated, decision-useful simulator for execution operations.

1) Modeling goal (don’t simulate for simulation’s sake)

Define the exact production decision first.

Typical decision targets:

1. Tactic selector (TWAP / POV / opportunistic passive / aggression burst)
2. Participation cap policy by regime
3. Kill-switch thresholds (spread blowout, cancel-rate shock, markout tail)
4. Inventory urgency policy under time-to-close constraints

If the simulator output does not map to one of the above, trim scope.

2) Minimal agent set (practical, not academic maximalism)

Start with 4 agent families:

1. Liquidity providers (LPs)
   • Quote both sides with inventory-aware skew
   • Pull quotes under toxicity/volatility stress
2. Noise traders
   • Randomized child orders with intraday intensity curve
3. Informed/metaorder flow
   • Directional child-order streams with persistence
4. Your execution agent
   • Candidate policies to evaluate (A/B/C)

This set is usually enough to recreate spread/queue/cancel dynamics relevant to slippage control.

3) State variables and observables

Track simulator state at event time:

• L1/L2 depth by side and level
• queue age/position distributions
• cancel/add/trade event intensities
• short-horizon markout (1s/5s/30s)
• spread, microprice drift, realized volatility
• realized slippage decomposition:
  • delay cost
  • impact cost
  • adverse selection
  • opportunity cost (missed fill)

Use the same schema as production TCA so model-vs-live comparison is frictionless.

4) Calibration pipeline (offline)

4.1 Stylized-fact constraints

Before fitting parameters, enforce baseline realism:

• spread distribution by time bucket
• depth imbalance distribution
• heavy-tail trade size distribution
• self-exciting event clustering (Hawkes-like burstiness)
• volatility–spread coupling

If these fail, optimization can overfit target KPI while violating market physics.

4.2 Objective function

Calibrate with a weighted distance over diagnostics:

[ \min_\theta ; \sum_k w_k d_k\big(\text{sim}_k(\theta), \text{real}_k\big) ]

Where diagnostics (k):

• level-wise depth profile
• cancel-to-trade ratio
• queue depletion half-life
• markout curve shape
• conditional slippage under participation buckets

Practical distance choices:

• Wasserstein distance for distributions
• MSE for curves
• tail-aware penalty for 95/99p slippage mismatch

4.3 Search strategy

Use staged search:

1. coarse global search (Sobol / random)
2. Bayesian optimization over top region
3. local refinement (CMA-ES or Nelder–Mead)

Parallelize by day/session regime (open, lunch, close, event day).

5) Regime conditioning

A single parameter set is fragile. Calibrate regime-conditioned parameter packs:

• calm: tight spread, stable replenishment
• fragile: high cancel propensity, shallow refill
• panic: spread jump + LP withdrawal + informed-flow burst

Regime labelers can be built from live observables:

• spread z-score
• cancel/add ratio surge
• top-of-book depth collapse rate
• short-window return variance

Execution policy should consume regime probabilities, not hard labels.

6) Stress scenario library (the core value)

Build replayable stress templates:

1. Cancel storm: synchronized LP quote pullback
2. One-sided toxicity burst: directional child-flow shock
3. Latency pocket: delayed acknowledgements + stale routing
4. Refill failure: depth recovery half-life doubled
5. Auction transition shock: pre-close imbalance ramp

Each scenario must report:

• expected shortfall of slippage (ES95)
• completion risk (% parent unfilled by horizon)
• worst 5-min drawdown in execution quality
• policy ranking stability (how often best policy flips)

7) Policy evaluation design

Compare execution policies on both mean and tail:

• mean implementation shortfall (IS)
• 90/95/99p slippage
• CVaR95 slippage
• missed-fill penalty
• post-trade markout

Recommended composite score:

[ \text{Score} = \mathbb{E}[IS] + \lambda_1 \text{CVaR}_{95}(IS) + \lambda_2 \cdot \text{UnfilledPenalty} ]

Tune (\lambda) from business constraints (urgency, inventory risk, mandate).

8) Online alignment loop (prevent simulator drift)

Simulator quality decays as market microstructure shifts. Add a live loop:

1. Daily compare live vs simulated diagnostics
2. Raise drift alerts when thresholds breach:
   • spread KS-stat
   • cancel ratio error
   • 95p slippage gap
3. Recalibrate parameter subset (not whole model) for stability
4. Promote only after shadow validation

Use champion–challenger governance:

• champion = current simulator params
• challenger = recalibrated params
• promote when challenger reduces weighted error over N sessions

9) Guardrails to avoid false confidence

• Do not optimize only average IS; tails are where blowups happen.
• Do not train/evaluate on same event regimes; enforce regime holdout.
• Do not hide latency assumptions; include gateway/exchange jitter models.
• Do not collapse all symbols; liquidity archetypes differ materially.
• Do not skip cost-of-control; conservative tactics can underfill badly.

10) Production adoption checklist

• Shared schema between simulator and real TCA
• Regime-conditioned parameter packs with versioning
• Scenario library with deterministic seeds
• Tail metrics (CVaR/ES) in acceptance gate
• Drift dashboard + recalibration trigger policy
• Champion–challenger promotion runbook

Quick implementation blueprint

• Data layer: event-level LOB + order lifecycle + latency stamps
• Sim core: event-driven engine (priority queue), pluggable agent policies
• Calibration service: distributed parameter search + diagnostics store
• Evaluation: policy sweep runner + scenario matrix + report exporter
• Ops: daily drift monitor, weekly recalibration window

Bottom line

A calibrated agent-based LOB simulator is not just a research toy. It is an execution risk wind tunnel: you test policies against nonlinear liquidity feedback before real capital pays the tuition. The edge is not perfect realism; the edge is reliable policy ranking under stress.