Fill-Hazard + Markout Coupled Slippage Controller Playbook

Date: 2026-02-25
Category: research (quant execution)

TL;DR

Most execution stacks model impact and fill probability separately, then wonder why live slippage tails explode.
Use a coupled controller: estimate (1) queue fill hazard, (2) adverse markout risk after fill, and (3) delay/opportunity cost if not filled, then choose passive vs aggressive action under a shared tail-budget.

1) Why this matters in production

For a buy order, passive quotes can look cheap in backtests but become expensive when:

queue priority is poor (low fill odds)
fill arrives only after microprice drifts up
non-fill forces urgent catch-up later at worse prices

If we optimize for spread capture alone, we buy false comfort.

2) Core decomposition

At decision time (t), for action (a_t):

[ \mathbb{E}[C_t(a_t)] = C^{impact}_t(a_t) + p^{fill}_t(a_t),C^{markout}_t(a_t) + (1-p^{fill}_t(a_t)),C^{delay}_t(a_t) ]

Where:

(C^{impact}): immediate crossing/temporary impact cost
(p^{fill}): fill probability over horizon (\Delta)
(C^{markout}): expected post-fill adverse movement (e.g., 5s/30s)
(C^{delay}): expected cost of waiting and retrying later

Risk-aware execution objective:

[ \min_{a_t}; \mathbb{E}[C_t(a_t)] + \lambda,\mathrm{Var}(C_t(a_t)) \quad \text{s.t.} \quad \Pr(C_t(a_t)>B_t) \le \delta_t ]

(B_t): remaining tail-budget slice
(\delta_t): tolerated breach probability

3) Modeling block A — Fill hazard

Model queue fill as survival/hazard process:

[ h_t = \Pr(\text{fill in }[t,t+dt] \mid \text{not yet filled}, x_t) ]

Useful features:

queue position estimate (shares ahead)
same-side cancel rate (queue decay)
opposite-side market order intensity (consuming flow)
spread state, top-book depth, depth imbalance
venue-specific latency + acknowledgement lag

Practical implementation:

piecewise-exponential hazard model or discrete-time logistic hazard
recalibrate per symbol bucket (ADV, spread regime, tick size)
maintain online correction factor to absorb intraday drift

4) Modeling block B — Adverse markout

Estimate expected adverse move conditional on fill:

[ M_{t,h} = \mathbb{E}[\text{sign}(side)\cdot (mid_{t+h}-fill_t) \mid \text{fill at }t, x_t] ]

Horizon set (h): 5s / 30s / 120s (or strategy-specific).

Feature candidates:

microprice-minus-mid dislocation
short-term signed trade imbalance
cancel burst intensity (fragility)
spread widening speed
local volatility shock flag

Heavy-tail handling:

Student-t residuals or quantile models (q50/q90/q95)
monitor calibration error by regime, not only global RMSE

5) Modeling block C — Delay/opportunity cost

When passive order misses, the parent schedule may become urgent.

Approximate delay cost:

[ C^{delay}_t \approx \alpha_t,\Delta t + \beta_t,\Delta POV + \gamma_t,\sigma_t\sqrt{\Delta t} ]

Interpretation:

(\alpha_t): alpha decay (edge leakage)
(\beta_t): convexity penalty for later catch-up aggression
(\gamma_t): volatility-driven uncertainty tax

This prevents the classic error: “passive now, panic later”.

6) Decision policy (passive ↔ aggressive switch)

Compute two candidate costs each cycle:

(J^{passive}_t): place/join queue
(J^{aggr}_t): cross spread now (or partial cross)

Switch if:

[ J^{passive}_t - J^{aggr}_t > \theta_t ]

Dynamic threshold (\theta_t) should tighten when:

remaining time-to-close is short
tail-budget burn rate is high
market enters fragility/toxicity regime

Recommended action space (small but effective):

Passive join (price = best bid/ask)
Passive improve by 1 tick (limited)
Partial marketable + residual passive
Full marketable slice with hard cap

7) Regime state machine

Use a 4-state controller:

Green: normal liquidity, stable markout
Yellow: rising toxicity or lower fill hazard
Orange: fragile queue + widening spread
Red: tail risk breach zone

State transitions from composite signal:

hazard deterioration score
markout q90/q95 jump
delay-cost acceleration
residual slippage budget burn

In Red:

shrink child size
restrict passive exposure duration
tighten per-order max slippage guard
allow temporary strategy throttle/pause

8) Vellab/KIS deployment blueprint

Data plane

KIS quote/trade stream + order ack/fill logs
local feature stream at 1–5s cadence
event-time alignment (exchange timestamp priority)

Model services

fill-hazard-service
markout-forecast-service
execution-policy-service
risk-sentinel (tail-budget + kill-switch ladder)

Storage / observability

append-only event log for replay (fills, cancels, decisions)
rolling calibration report (coverage, bias, drift)
regime timeline dashboard (Green→Red transitions)

Must-have guards

stale market data gate (fallback mode)
missing ack/fill reconciliation watchdog
per-symbol and portfolio-level throttle caps
auto-downgrade when model confidence collapses

9) Validation protocol (before real capital scaling)

Do not ship from average bps improvements alone.

Track at least:

mean/median slippage
q90/q95/q99 slippage
fill ratio + underfill ratio
catch-up aggression frequency
post-fill markout distribution by state
tail-budget breach frequency

A/B setup:

Baseline: existing impact-only or static policy
Candidate: coupled hazard+markout controller
Same universe, same notional buckets, same hard risk limits

Success condition (practical):

q95 down, breach frequency down
no material underfill blow-up
stable performance across open/lunch/close regimes

10) Failure modes to watch

Queue position mismeasurement → fake fill probabilities
Label leakage in markout model → backtest mirage
No latency accounting → stale control actions
Tail budget ignored near close → forced panic execution
Overcomplex policy → hard-to-debug behavior under stress

Keep policy interpretable; complexity should be in modeling, not hidden branching logic.

11) 10-day rollout plan

Day 1–3: shadow mode (predict only, no control)
Day 4–6: tiny-notional pilot + strict kill-switch rails
Day 7–8: expand symbols in liquid buckets only
Day 9–10: evaluate q95 + underfill + breach KPIs; gate scale-up

References (starting points)

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.
Taranto, D. et al. (2016/2018), Propagator-family linear impact models.
Stoikov, S. (2017), The Micro-Price: A High Frequency Estimator of Future Prices.

The practical edge is coupling: fill hazard + markout + delay cost in one control loop. That is where slippage tails usually hide.