Drop-Copy Lag & Phantom Residual Slippage Playbook

Why this matters

Some execution stacks treat drop-copy as the source of truth for fills/positions during live slicing. When drop-copy lags (or arrives out of order), the router can hallucinate residual quantity that is already filled, then submit extra child orders.

That creates a hidden slippage loop:

1. perceived underfill ->
2. unnecessary catch-up aggression ->
3. overfill unwind / hedge churn ->
4. extra spread + impact + fee drag.

This is not classic market impact. It is a state-consistency tax.

Failure pattern (execution branch view)

At each decision point for parent order P:

• Q_true: actual remaining quantity
• Q_seen: remaining quantity inferred from delayed drop-copy
• delta = Q_seen - Q_true

If delta > 0, strategy believes it is underfilled and over-trades.

Expected incremental cost from hallucinated residual:

E[ExtraCost] = E[ delta * (spread + impact + urgency_penalty) ] + E[unwind_cost]

The worst tails appear near deadlines (close, auction cutoff, schedule end), where urgency multipliers are steep.

Core metrics

1) Drop-Copy Delay Quantile (DCDQ)

DCDQ_p95 = p95( t_dropcopy_fill - t_true_fill )

Track by venue, symbol-liquidity bucket, and time-of-day regime.

2) Phantom Residual Ratio (PRR)

PRR = sum(max(Q_seen - Q_true, 0)) / sum(parent_qty)

Measures how much parent size was falsely considered "remaining".

3) Catch-up Overshoot Ratio (COR)

COR = sum(overfill_qty_due_to_state_lag) / sum(parent_qty)

Quantifies over-trading induced by stale execution state.

4) Reconciliation Shock Cost (RSC)

RSC = unwind_cost + forced_hedge_adjustment_cost

Use arrival-to-unwind and markout-adjusted accounting.

5) State Divergence Half-life (SDH)

Median time from divergence detection (Q_seen != Q_true) to restored consistency.

Modeling framework

A) Two-layer state model

• Latent true state from exchange/order-gateway ACK + execution reports
• Observed state from drop-copy stream used by strategy

Model P(divergence | latency, venue, load, message_rate, reconnect_state) and E[cost | divergence].

B) Tail-first objective

Do not optimize only mean bps.

Use:

J = E[slippage] + lambda1 * q95(slippage) + lambda2 * RSC

Because divergence events are sparse but expensive.

C) Deadline interaction term

Include time_to_deadline interaction:

cost ~ divergence * f(time_to_deadline)

This captures convex urgency tax when stale state persists late.

Execution controller (state machine)

STATE 1: SYNCED

Criteria:

• DCDQ p95 below threshold
• PRR low and stable

Policy:

• Normal slicing logic
• Full tactic set allowed

STATE 2: UNCERTAIN

Criteria:

• DCDQ rising, intermittent divergence

Policy:

• Reduce max child size
• Increase minimum inter-dispatch gap
• Require dual confirmation for aggressive catch-up

STATE 3: DEGRADED

Criteria:

• Persistent divergence or reconnect burst
• PRR/COR breach

Policy:

• Freeze non-essential repricing churn
• Prefer queue-preserving amend over cancel/replace
• Cap urgency escalation slope
• Enforce reconciliation window before large sweeps

STATE 4: SAFE_RECONCILE

Criteria:

• Severe divergence + deadline stress

Policy:

• Stop autonomous catch-up
• Run deterministic reconciliation snapshot
• Resume only after state lock restored

Practical guardrails

1. Confidence-weighted residual

   • Use Q_effective = w * Q_seen + (1-w) * Q_conservative
   • Decrease w as DCDQ/PRR worsens.

2. Dual-feed sanity checks

   • Compare gateway-native execution view vs drop-copy view.
   • Trigger degraded state if disagreement persists beyond tolerance.

3. Reconciliation cadence by regime

   • Tighten cadence around open/close/auction windows.

4. Replayable incident packet

   • Persist event timeline: send, ACK, fill, drop-copy arrival, decision.
   • Needed for root-cause attribution and model recalibration.

5. No blind urgency under uncertainty

   • If state confidence falls, urgency multiplier must saturate (hard cap).

Validation plan

Offline

• Reconstruct true-vs-seen residual timelines from historical packets.
• Counterfactual replay: baseline controller vs divergence-aware controller.
• Compare mean, q95, q99 slippage and RSC.

Shadow live

• Run divergence-aware policy in observe-only mode.
• Emit would-have-acted decisions and projected cost deltas.

Canary

• Start with low ADV symbols and low participation caps.
• Auto-rollback if completion shortfall or q95 cost breaches policy bands.

Operator checklist

• DCDQ monitored per venue/session bucket
• PRR/COR/RSC dashboards with alert thresholds
• State machine transitions audited and explainable
• SAFE_RECONCILE tested in drills (not only incidents)
• Weekly recalibration includes divergence features

Bottom line

When fill state is delayed, execution can pay a phantom residual tax that looks like random market noise. Treat drop-copy lag as a first-class risk factor, model it explicitly, and bind urgency to state confidence.

That is how you reduce tail slippage without sacrificing completion discipline.