Regime-Switching Slippage Surface (Participation × Duration × Queue-Risk) Playbook

Date: 2026-02-26
Category: Research (Execution / Slippage Modeling)
Scope: Intraday execution for single-name + basket routing (KR-first, global portable)

Why this model exists

Most slippage models fail for practical reasons:

• Static coefficients break when market regime flips (open/close, news, VI-adjacent stress).
• One-factor assumptions (only participation or only duration) miss nonlinear interaction.
• Queue-risk blindness underestimates cost when cancel pressure spikes and passive fills vanish.

A practical execution stack needs a model that answers one question continuously:

"Given current participation, remaining duration, and queue toxicity, what is expected slippage in this regime, and what action keeps tail risk bounded?"

This playbook builds exactly that using a regime-switching slippage surface.

Core concept

Model expected slippage in basis points as:

[ \mathbb{E}[IS_t \mid x_t] = \sum_{k=1}^{K} p_t^{(k)} \cdot g_k(x_t) ]

• (x_t): feature vector (participation, duration, spread, volatility, queue signals, venue state)
• (g_k): regime-specific slippage surface for regime (k)
• (p_t^{(k)}): online probability that regime (k) is active

Recommended regimes ((K=3)):

1. Normal liquidity
2. Fragile liquidity (book looks thick, but replenishment weak)
3. Stress / dislocation (high cancel bursts, spread jumps, queue churn)

This avoids the "single average model" trap and gives direct control levers.

Feature design (minimal but production-strong)

A) Execution geometry

• Participation rate (\pi_t = q_t / V_t)
• Remaining horizon (T_t) (seconds/minutes)
• Urgency ratio (u_t = q_{remain} / T_t)

B) Market state

• Instant spread (ticks/bps)
• Short-horizon realized vol (e.g., 30s, 2m)
• Depth slope around top-of-book
• Auction / session flags (open, lunch, close, auction window)

C) Queue-risk / toxicity

• Cancel-to-trade ratio (rolling)
• Best-level depletion velocity
• Queue age decay (our posted order survival)
• Microprice drift vs mid-price drift divergence

D) Venue/routing context

• Venue id (KRX/NXT/alt)
• Venue-local fill hazard
• Cross-venue top-of-book divergence

Slippage surface form

Use a flexible but inspectable form per regime:

[ g_k(x) = \beta_{0k}

• \beta_{1k},\sigma\sqrt{\pi}
• \beta_{2k},\sigma\sqrt{T}
• \beta_{3k},\text{spread}
• \beta_{4k},\text{queueTox}
• \beta_{5k},\pi\cdot\text{queueTox}
• \beta_{6k},\sqrt{\pi}\cdot\sqrt{T} ]

Notes:

• (\sqrt{\pi}) term aligns with concave impact stylized fact (square-root behavior).
• (\sqrt{T}) term captures duration/time-risk exposure.
• Interaction terms capture where real losses happen: "high participation + toxic queue".

If you need extra flexibility, replace with monotonic GBM per regime (with constraints on (\pi), spread, toxicity).

Regime inference (online)

Use a lightweight Hidden Markov Model (or switching state-space filter):

1. Compute residual: [ r_t = IS_t - g_{z_t}(x_t) ]
2. Update regime posterior (p_t^{(k)}) using transition matrix (A) and emission likelihood.
3. Apply persistence prior (avoid micro-flipping every second).

Practical transition prior:

• High persistence on diagonal (0.92–0.98)
• Faster transition into stress than out of stress

This lets controller react early without overtrading on noise.

Training pipeline

1) Label construction

For each child fill or micro-batch:

• benchmark (arrival/decision)
• realized slippage (bps)
• synchronized microstructure snapshot at send time

2) Data split

• Purged time split by day/session
• Symbol-group split to test cross-name generalization

3) Fit baseline surface

• Start with pooled model (all data)
• Initialize regime components via residual clustering (e.g., k-means on residual + toxicity + spread jump)

4) Fit switching model

• EM optimization for (g_k), transition matrix, emission variance
• Regularize to prevent degenerate regimes

5) Calibrate tails

• Add quantile layer (q90/q95) per regime for risk budget control

What to optimize (objective)

For production, optimize risk-adjusted execution cost, not mean IS only:

[ J = \mathbb{E}[IS] + \lambda_1,\text{CVaR}_{95}(IS) + \lambda_2,\text{UnderfillPenalty} ]

Why:

• Mean improves can hide rare but expensive blowups.
• Underfill penalty keeps model from becoming too passive during stress.

Controller integration (how model actually saves money)

At each control step:

1. Estimate expected slippage + tail for candidate actions:
   • keep pace
   • slow participation
   • speed up to reduce time-risk
   • passive-heavy vs aggressive-heavy mix
   • venue rebalance
2. Pick action minimizing projected objective under current budget.

Example policy guardrails

• If (P(\text{stress}) > 0.65) and queue toxicity rising:
  • reduce passive dwell time
  • tighten stale quote cancel threshold
  • cap single child notional
• If (P(\text{normal}) > 0.8) and spread tight:
  • restore passive participation
  • lengthen quote life to save spread cost

Evaluation scorecard

Modeling quality

• MAE / RMSE of IS (overall + by regime)
• Calibration error of regime probabilities
• Quantile coverage error (q90/q95)

Execution impact

• Delta mean IS vs baseline schedule
• Delta q95/q99 IS (must improve)
• Completion rate and underfill slippage transfer

Stability

• Regime flip rate per hour (too high = unstable)
• Action churn (cancel/replace inflation)
• Venue oscillation frequency

KR market implementation notes

• Separate models for opening call/closing auction windows; don’t force one continuous model.
• Treat VI-adjacent states as stress prior bump even before full trigger.
• Preserve conservative fallback schedule if feed latency or queue features go stale.
• Maintain venue-local toxicity estimate; stress can be asymmetric across venues.

Failure modes and protections

1. Regime overfitting

   • Symptom: excellent backtest, unstable live posteriors
   • Fix: stronger transition prior + fewer regimes

2. Hidden latency bias

   • Symptom: model underestimates costs during burst periods
   • Fix: add feature freshness lag and gateway RTT features

3. Overreaction controller

   • Symptom: action churn increases fees without slippage gain
   • Fix: hysteresis + minimum hold time per mode

4. Data leakage via post-trade features

   • Symptom: impossible offline accuracy
   • Fix: enforce strict event-time cutoff at order send

Practical rollout plan (2 weeks)

Week 1

• Build unified feature table and IS label stream
• Train 1-regime baseline + 3-regime switching prototype
• Backtest with replay and tail metrics

Week 2

• Shadow mode in paper/live-sim routing
• Compare decisions vs current controller
• Enable only low-risk levers first (quote dwell, child cap)
• Promote to active once q95 improvement is stable for 5+ sessions

Reference pointers

• Square-root impact stylized facts and propagator framing (metaorder impact literature; practical summaries include Emergent Mind overview, 2025 update).
• Participation-vs-duration modeling debate and the operational need to separate physical impact vs time risk (institutional execution literature; Talos practitioner write-up, 2025).
• Regime-switching/state-space methods from time-series econometrics adapted to execution control.

TL;DR

Use a regime-switching slippage surface instead of one static model.

Predict slippage as a mixture of regime-specific surfaces over participation, duration, and queue toxicity; then route/control with tail-aware objective.

You’ll usually get modest mean improvement, but the real win is material q95/q99 slippage drawdown control when market microstructure turns hostile.