Cross-Venue ACK Dispersion & Residual-Estimation Slippage Playbook

Why this matters

Most execution logic assumes your remaining parent size is known almost exactly in real time.

That assumption breaks in fragmented markets:

• child orders are sent to multiple venues,
• acknowledgements/fills arrive with different latencies,
• cancel confirms arrive even later,
• router state can be temporarily wrong about true residual.

When residual state is stale, the controller makes expensive choices:

1. Under-estimates residual → slows down too early, then pays late catch-up convexity.
2. Over-estimates residual → keeps pressing and risks overfill/self-competition churn.
3. Ping-pongs urgency on noisy ACK updates, paying spread and queue-reset tax.

Execution slippage is then driven by state-estimation lag, not only market impact.

Core failure mechanism

Let:

• R_true(t): true remaining parent quantity,
• R_est(t): router-estimated remaining quantity,
• e_R(t) = R_est(t) - R_true(t): residual estimation error.

If ACK/fill/cancel events are delayed and heterogeneous across venues, e_R(t) becomes regime-dependent.

A simple dispersion driver:

D_ack(t) = p90(ack_latency_v,t) - p10(ack_latency_v,t)

for active venues v.

As D_ack grows, event-ordering ambiguity rises; residual control acts on stale state and branch losses become nonlinear near deadline windows.

Branch-cost decomposition

Expected cost can be decomposed as:

E[Cost] = C_base + C_underfill_catchup + C_overfill_unwind + C_churn + C_toxicity_miss

Where:

• C_base: normal spread/impact/fees under synchronized state,
• C_underfill_catchup: late aggressive completion when residual was under-estimated,
• C_overfill_unwind: liquidation/hedge correction cost after over-execution,
• C_churn: cancel/replace and queue-reset costs from oscillating urgency,
• C_toxicity_miss: wrong passive/aggressive mix chosen due to stale residual confidence.

The hidden tax is convex because mis-estimation interacts with time pressure.

Metric stack

1) ACK Dispersion Spread (ADS)

ADS = p90_ack_latency - p10_ack_latency across active venues.

• High ADS means asynchronous state truth.

2) Residual Error Band (REB)

Posterior interval width for residual estimate, e.g. q90(R_true|events) - q10(...).

• Wide REB means controller should reduce brittle one-shot urgency shifts.

3) Deadline Catch-up Convexity (DCC)

Marginal expected cost per residual unit as time-to-deadline shrinks.

• Rising DCC means residual uncertainty is becoming expensive fast.

4) Overfill Pressure Index (OPI)

Probability-weighted overfill risk under current in-flight child orders and cancel-lag.

• High OPI means continued aggression can create unwind drag.

5) ACK Ordering Violation Rate (AOVR)

Rate of event sequences implying ambiguous or inverted local ordering for state updates.

• Elevated AOVR means event ingest rules are too naive for current latency regime.

Control policy state machine

STATE 1 — SYNCHRONIZED

Conditions:

• low ADS,
• narrow REB,
• low AOVR.

Policy:

• standard residual-tracking controller,
• normal venue allocation and urgency schedule.

STATE 2 — DESYNC_WATCH

Conditions:

• ADS/REB rising,
• occasional ordering violations,
• moderate DCC.

Policy:

• downweight fast urgency flips,
• increase residual confidence threshold before large aggression jumps,
• cap simultaneous in-flight child notional.

STATE 3 — DESYNC_ACTIVE

Conditions:

• persistent high ADS/AOVR,
• wide REB,
• rising OPI or DCC.

Policy:

• switch to uncertainty-aware residual posterior control,
• prioritize venues with stable ACK behavior,
• constrain cancel/reprice churn and avoid oscillatory response.

STATE 4 — SAFE

Conditions:

• residual confidence collapse near deadline,
• combined overfill + catch-up risk exceeds budget.

Policy:

• halt non-essential micro-updates,
• execute pre-approved completion ladder with bounded overfill risk,
• prefer reliability over local spread optics.

Use asymmetric hysteresis to avoid state flapping.

Modeling pattern (production)

1. Residual posterior instead of point estimate

   • maintain P(R_true | event stream, per-venue latency model).
   • propagate uncertainty to tactic scoring, not just dashboarding.

2. Per-venue ACK latency regime model

   • online estimates by venue × session segment × stress state,
   • detect sudden latency skew shifts (not only mean drift).

3. Uncertainty-aware action objective

   • optimize mean + q95 cost with explicit overfill and deadline penalties,
   • penalize tactic sensitivity to residual uncertainty.

4. Event-order robust ingestion contract

   • idempotent apply rules,
   • monotonic sequence/causality guards where available,
   • reconciliation path for delayed cancel/fill confirms.

5. Counterfactual replay with latency perturbations

   • inject ACK-latency shocks and ordering permutations into historical runs,
   • verify state transitions and tail-cost protection before production promotion.

Practical rollout checklist

• Implement ADS/REB/OPI/AOVR dashboards by strategy and venue bucket.
• Add DESYNC state machine with manual override + incident logging.
• Limit in-flight exposure as a function of REB and DCC.
• Canary in one strategy with strict q95 and overfill guardrails.
• Weekly attribution: split slippage into synchronized vs desynchronized windows.

Bottom line

In fragmented execution, residual size is not a constant truth—it is a latency-conditioned belief state.

If your router acts on stale residual estimates as if they were exact, you quietly pay slippage through late catch-up, overfill unwind, and churn. Model ACK dispersion explicitly and let uncertainty drive control logic, especially when deadline convexity turns small state errors into expensive outcomes.