Marketable-Limit Protection-Band Oscillation Slippage Playbook

Date: 2026-03-11
Category: research (quant execution / slippage modeling)

Why this matters

Many desks use marketable-limit orders with protection offsets to avoid catastrophic prints.
That guardrail is necessary, but under fast microstructure changes it can create a hidden loop:

1. order is sent with a protection band,
2. price touches band and rejects/underfills,
3. strategy immediately widens/reprices,
4. next attempt lands after microprice drift,
5. repeated loops convert safety logic into convex slippage.

So the problem is not "limit vs market."
It is control-loop stability under moving reference prices.

Core concept: Protection-Band Oscillation Tax (PBOT)

Decompose realized cost impact from protection logic as:

[ \text{PBOT} = \text{BandRejectCost} + \text{RepriceDelayDrift} + \text{RetryQueueResetTax} - \text{TailLossAvoided} ]

Execution objective:

[ \min_a; \mathbb{E}[C|a] + \lambda,\mathrm{CVaR}_{95}(C|a) + \eta,\mathrm{DeadlineMiss}(a) ]

where action (a) controls:

• protection offset width,
• reprice step size,
• retry cooldown,
• passive/aggressive branch mix,
• completion fallback thresholds.

Key insight: minimizing mean cost without loop-stability constraints often increases q95/q99 slippage.

Detection signals (oscillation observability)

Construct a Band Oscillation Score (BOS) from the following:

1. Band-Touch Reject Rate (BTR)
   Fraction of attempts that reach protection boundary and fail to complete.

2. Reprice Loop Frequency (RLF)
   Number of cancel-reprice cycles per parent-minute or per 1% ADV child schedule bucket.

3. Anchor Staleness Gap (ASG)
   Time delta between anchor quote snapshot and actual exchange interaction.

4. Miss-to-Chase Slope (MCS)
   Marginal cost increase from each additional retry cycle.

5. Protection Utilization Ratio (PUR)
   Distance consumed inside protection band before fill; persistent high PUR implies chronic under-sizing.

6. Partial-Fill Fragmentation Index (PFI)
   Parent completion fragmentation caused by repeated micro underfills.

State mapping example:

• NORMAL: low BTR, low RLF, stable ASG
• TIGHT: moderate rejects, rising PUR, still controllable
• OSCILLATING: high RLF + positive MCS, retries driving cost acceleration
• SAFE: tail or deadline budget breach; switch to completion-preserving policy

Use hysteresis and minimum dwell time to avoid rapid state flapping.

Modeling stack

Layer 1 — Child-attempt branch model

For each child attempt, model:

• full fill / partial fill / reject probabilities,
• expected delay to next actionable attempt,
• conditional post-fill markout,
• queue reset penalty after cancel-reprice.

Targets:

• mean + q90/q95 attempt cost,
• expected residual after attempt,
• probability of entering oscillation state after attempt.

Layer 2 — Parent-episode dynamics

Aggregate attempts into parent-order dynamics:

• completion-time distribution,
• cumulative oscillation tax,
• tail-cost concentration by state,
• deadline breach probability under policy choice.

This prevents "attempt looked fine" bias when episode-level convexity is the true risk.

Control policy design

1. State-aware protection width
   In NORMAL/TIGHT, keep narrow offsets for price discipline.
   In OSCILLATING, widen using bounded schedule to reduce reject loops.

2. Fresh-anchor gating
   Recompute protection anchor when ASG exceeds threshold; block reuse of stale reference prices.

3. Monotone retry policy
   Enforce bounded monotone aggression (no ping-pong narrowing/widening) to reduce control oscillation.

4. Retry cooldown floor
   Add minimum inter-attempt delay when RLF spikes; avoid self-induced queue churn.

5. Partial-preservation branch
   If PFI rises, prioritize completing larger residual chunks over micro-optimization of each child.

6. Tail-budget circuit breaker
   If projected q95 burn exceeds budget, shift to SAFE profile with completion-first constraints.

Data contract (minimum)

Per attempt:

• decision/send/gateway/ack/reject/fill timestamps,
• decision-time and interaction-time top-of-book + depth snapshot,
• protection offset parameters and anchor timestamp,
• cancel reason and reprice step metadata,
• attempt index + linkage to parent chain,
• markout ladder (e.g., 100ms/500ms/1s/5s),
• residual size, urgency bucket, deadline metadata.

Without precise attempt linkage and anchor freshness tracking, PBOT attribution collapses.

Calibration & monitoring loop

Weekly

• Refit branch probabilities by liquidity regime and session segment.
• Re-estimate BOS thresholds and transition matrix.
• Validate q95 calibration by state and urgency tier.

Daily

Monitor:

• PBOT decomposition (reject vs delay drift vs queue-reset tax),
• state occupancy and persistence,
• retry-count distribution per parent,
• completion ratio at fixed deadline budgets.

Intraday guards

• BOS above threshold for N windows -> escalate to OSCILLATING controls.
• MCS slope > threshold -> cap retry count + widen/cross fallback.
• q95 forecast breach + completion stress -> SAFE fallback.

Rollout plan

1. Shadow mode (1–2 weeks): compute BOS/PBOT and simulate policy switches.
2. Canary notional (5–10%): enable fresh-anchor gating + retry cooldown.
3. Scale-up: activate state-aware protection width and monotone retry policy.
4. Guarded expansion: monitor deadline hit-rate and tail-cost drift before full rollout.

Rollback triggers:

• completion hit-rate deterioration,
• BOS instability (frequent unsafe flapping),
• q95 cost expansion beyond tolerance.

Common failure modes

• Treating protection-band rejects as harmless "safe misses" instead of delayed-cost transfer.
• Repricing from stale anchors and compounding drift.
• Optimizing average fill price while ignoring episode-level retry convexity.
• Allowing bi-directional parameter swings (tighten/widen/tighten) that create controller oscillation.

Bottom line

Protection bands are essential, but they are also a feedback controller.
When markets move quickly, unstable retry/reprice loops can dominate slippage.

Model band reject dynamics + retry convexity explicitly, detect oscillation early, and switch to stable state-aware policies before tail costs compound.