Queue-Position-Aware Fill + Slippage Coupled Model Playbook

Date: 2026-02-24 (KST)

TL;DR

Most slippage models treat cost and fill probability as separate problems. In live execution they are tightly coupled:

• if you chase fills aggressively, impact cost rises
• if you protect price too much, fill risk rises and later urgency explodes cost

This playbook proposes a coupled model that jointly tracks:

1. Queue position / queue survival (will my passive slice fill?)
2. Conditional slippage under fallback actions (cross/step-up/repost)
3. Policy switch thresholds from remaining slippage budget and time-to-close

Goal: reduce p95 implementation shortfall while preserving completion SLA.

1) Why the Coupled View Matters

Common failure pattern in production:

1. passive posting model says “good quote, high edge”
2. queue gets depleted/repriced/canceled against us
3. late-stage catch-up switches to aggressive taking
4. final shortfall > model forecast due to urgency convexity

Root cause: the model predicted fill and slippage on disconnected assumptions.

2) Problem Setup

For each child decision at time (t):

• action (a_t \in {\text{post},\text{join},\text{step-up},\text{take}})
• horizon (\Delta) (e.g., 1s, 5s, 15s)
• target metrics:
  • fill probability (P(F_{t,\Delta}=1 \mid x_t, a_t))
  • conditional slippage quantiles (Q_q(S_{t,\Delta} \mid x_t, a_t, F))

State (x_t) should include:

• queue position estimate (shares ahead, level aging)
• cancel/add/trade intensities near touch
• spread state, microprice drift, imbalance/OFI
• latency + reject bursts (infra confounders)
• parent-order urgency (remaining qty, time, budget headroom)

3) Core Modeling Architecture

3.1 Queue Survival / Hazard Layer

Model queue departure hazard for our resting order:

[ \lambda_t = f_{haz}(x_t) ]

Then approximate fill survival:

[ P(F_{t,\Delta}=1) \approx 1-\exp\left(-\int_t^{t+\Delta} \lambda_u,du\right) ]

Practical note:

• start with piecewise-constant hazard by spread/depth bucket
• upgrade to gradient-boosted survival or discrete-time logistic hazard

3.2 Action-Conditional Slippage Layer

For each action, fit quantile models:

[ \hat S_q(x_t,a_t)=g_{a_t,q}(x_t) ]

• passive actions: include non-fill penalty via continuation value
• aggressive actions: include impact convexity vs participation and local depth

3.3 Continuation Value (Fill ↔ Future Cost Bridge)

Define expected future catch-up cost if not filled now:

[ C_{cont}(t)=E\left[S_{future}\mid \neg F_{t,\Delta},;u_{t+\Delta}\right] ]

Decision score:

[ J(a_t)=E[S_{t,\Delta}\mid a_t]+(1-P(F_{t,\Delta}\mid a_t))\cdot C_{cont}(t) ]

Pick action minimizing (J(a_t)) under completion constraints.

4) Live Control Logic (Desk-Friendly)

Use a 3-zone controller with hysteresis:

• Harvest: high fill odds + healthy budget -> favor passive/posting
• Balance: mixed signals -> blend join/step-up with capped aggression
• Salvage: low fill odds + tight deadline -> controlled taking with clip ladder

Trigger variables:

• remaining slippage budget headroom
• expected completion gap at horizon
• queue-fragility index (QFI)

Example headroom:

[ Headroom_t=B-(IS_{realized,t}+\hat IS^{remaining}_{95,t}) ]

If headroom < 0 and completion gap rising, escalate one zone.

5) Features That Actually Move the Needle

Minimal high-signal set:

1. shares ahead at best level (and its change rate)
2. cancel-to-add ratio near touch
3. trade-through frequency (short window)
4. microprice minus mid (signed)
5. short-horizon OFI z-score
6. spread regime + depth regime
7. local infra quality (ack latency, reject spikes)

Avoid feature bloat before timestamp hygiene is proven.

6) Validation Protocol

Do not validate on average bps only.

Required scorecard slices

• liquid / mid / thin symbols
• open / midday / close / event windows
• buy vs sell asymmetry
• queue-depth deciles

Required metrics

• calibration of fill probability (Brier + reliability curves)
• pinball loss for q50/q90/q95 slippage
• completion SLA hit rate
• tail exceedance magnitude (actual - q95 forecast)

Promotion rule: tail improvement must not be bought by unacceptable completion failure.

7) Rollout Plan (Practical)

• D1–D3 Shadow: compute (P(F)), (\hat S_q), (J(a)), no routing impact
• D4–D6 Assist: show recommendations, trader/engine keeps veto
• D7–D10 Guarded Auto: apply for low-risk bucket with hard kill-switches

Hard stops:

• q95 exceedance blowout > threshold for 2 sessions
• completion SLA breach above desk tolerance
• detected infra incidents contaminating market signals

8) Common Failure Modes

• queue position estimate drifts due to hidden/iceberg interactions
• fill model ignores cancel cascades (false optimism)
• non-fill continuation cost underestimated in late schedule
• no separation of infra latency events vs true toxicity
• controller flips too fast (missing dwell-time hysteresis)

9) Implementation Checklist

• strict decision-time feature contract
• queue hazard model with live calibration monitor
• action-conditional slippage quantile models
• continuation-value estimator for non-fill penalty
• 3-zone controller with hysteresis + kill-switches
• slice dashboard: fill calibration / q95 coverage / SLA
• post-trade attribution: model vs infra vs venue conditions

References

• Almgren, R., & Chriss, N. (2000). Optimal Execution of Portfolio Transactions.
• Huang, W., Lehalle, C.-A., & Rosenbaum, M. (2015). Simulating and analyzing order book data: The queue-reactive model. JASA. DOI: 10.1080/01621459.2014.982278
• Taranto, D. E., Bormetti, G., Bouchaud, J.-P., Lillo, F., & Toth, B. (2016). Linear models for the impact of order flow on prices I: Propagators. arXiv:1602.02735
• Stoikov, S. (2017). The Micro-Price: A High Frequency Estimator of Future Prices. SSRN 2970694
• Gatheral, J. (2010). No-Dynamic-Arbitrage and Market Impact. Quantitative Finance.

Closing Note

Execution quality is not a single “slippage prediction” number.

It is a sequential control problem where queue survival and impact are entangled. Model them together, and the desk stops paying urgency tax that was invisible in disconnected models.