Latency-Arbitrage Shadowing & Quote-Fade Slippage Playbook

Date: 2026-03-10
Category: research (quant execution / slippage modeling)

Why this matters

Many execution stacks still assume that if top-of-book looks good at decision time, passive joins or light aggression should be fine.

In practice, when fast participants detect stale or vulnerable quotes, they can shadow and fade before slower flow lands:

• your passive order misses as queue quality collapses,
• your aggressive take lands after microprice has already moved,
• repeated retries convert small edge into convex slippage.

This is not just generic "high volatility". It is a latency-arbitrage interaction regime with distinct timing signatures.

Core concept: Shadow-Fade Tax (SFT)

Model expected child-order cost as:

[ \text{SFT} = \text{StaleHitMarkout} + \text{FadeMissChaseTax} + \text{RetryTimingTax} - \text{SpreadCapture} ]

Execution objective:

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

where action (a) controls join/improve/take choice, price offsets, retry cooldown, and venue mix.

Detection signals (shadowing observability)

Build a Shadowing Pressure Score (SPS) from:

1. Quote Fade Lead (QFL)
   Median time between decision-time top quote and its cancellation/price shift before arrival.

2. Stale Interaction Rate (SIR)
   Fraction of child orders interacting with quotes that were top-of-book only briefly before adverse move.

3. Microprice Drift-to-Latency Ratio (MDLR)
   Microprice drift during order transit divided by end-to-end decision latency.

4. Retry Adverse Escalation (RAE)
   Incremental cost between first attempt and subsequent retries in short windows.

5. Queue Quality Collapse (QQC)
   Effective executable depth decay vs displayed depth under interaction pressure.

State mapping example:

• STABLE: low SPS, normal quote persistence
• SHADOWED: elevated fade lead + transit drift
• PREDATORY: high stale-hit and retry escalation
• SAFE: uncertainty or tail-budget breach; preserve completion integrity

Use hysteresis + minimum dwell time to avoid flapping.

Modeling stack

Layer 1 — Child branch model

Estimate branch outcomes:

• passive fill / partial / miss,
• aggressive fill with post-fill markout,
• retry branch timing and price-worsening trajectory.

Recommended targets:

• mean + q90/q95 slippage,
• branch probabilities,
• expected completion delay contribution.

Layer 2 — Episode model

Aggregate child outcomes into parent-order episode cost:

• first-attempt vs retry share,
• tail concentration by state,
• deadline-pressure amplification.

This prevents “average fill looked okay” blindness when tail losses concentrate in SHADOWED/PREDATORY windows.

Control policy design

1. Latency-aware aggression floor
   In SHADOWED/PREDATORY states, raise minimum aggression to avoid repeated stale touches.

2. Retry cooldown + cap
   Hard-limit rapid retries when RAE exceeds threshold; enforce cooldown before re-entry.

3. Dynamic quote validity window
   Tighten acceptable quote-age/validity gates as MDLR rises.

4. Venue resilience weighting
   Route toward venues with better realized quote survival under stress, not just headline spread.

5. Tail-budget kill switch
   If q95 burn rate breaches limit, move to SAFE profile (completion-first, lower path complexity).

Data contract (minimum)

Per child order:

• decision/send/gateway/ack/fill/cancel timestamps (high-fidelity),
• decision-time and interaction-time top-N book snapshots,
• queue position proxy features,
• venue/route hop metadata,
• markout ladder (e.g., 50ms/250ms/1s/5s),
• retry chain linkage (attempt index, inter-attempt delay),
• parent residual size + deadline metadata.

Without timestamp quality and attempt linkage, shadowing attribution collapses into noise.

Calibration & monitoring loop

Weekly

• Refit branch model by liquidity bucket and time-of-day.
• Re-estimate SPS thresholds and transition probabilities.
• Validate coverage of q95 forecasts by state.

Daily

Monitor:

• SFT decomposition (stale-hit vs miss-chase vs retry tax),
• state occupancy and transition frequency,
• completion ratio under elevated SPS,
• tail-cost concentration by venue and tactic.

Intraday guards

• SPS sustained above threshold for N windows → escalate policy.
• Retry tax above budget → cap retries and widen cooldown.
• q95 coverage failure + completion stress → SAFE fallback.

Rollout plan

1. Shadow mode (2 weeks): compute SPS and simulated actions only.
2. Canary (5–10% notional): enable retry controls + quote-validity gates.
3. Scale-up: add venue weighting and state-specific aggression floors.
4. Rollback triggers: completion drop, q95 blowout, or state instability.

Common failure modes

• Treating quote fade as random noise instead of adversarial timing structure.
• Optimizing spread capture while ignoring retry convexity.
• Measuring only average slippage and missing state-concentrated tails.
• Ignoring transport/decision latency decomposition, making controls untargetable.

Bottom line

When latency-arbitrage shadowing is active, slippage is driven less by displayed spread and more by who reaches the quote first.

Model stale-hit and fade-miss branches explicitly, then control retries and quote-validity dynamically. That is how you keep small timing disadvantages from compounding into large tail-cost leakage.