Order-Intent Inversion Slippage Playbook

Date: 2026-03-14
Category: research
Focus: Control-path asymmetry (cancel vs new) in live execution

1) Why this needs its own slippage model

Many routers assume this safe sequence:

1. Old child order is canceled
2. Cancel is confirmed
3. New child order is placed

In production, this often breaks. New orders may travel through a faster path than cancels (different gateways, signatures, risk hooks, sequencer queues, or throttles).

Result: intent inversion — the strategy "thinks" it replaced risk, but both old and new intentions are simultaneously live.

That creates hidden slippage even when spread/impact conditions look normal.

2) Hidden cost channels

When cancel-path latency > new-path latency, cost leaks through four branches:

1. Dual-Live Exposure Tax (DLET)
   Temporary over-exposure while old and new leaves coexist.

2. Self-Interaction Tax (SIT)
   Internal flow collision risk (or forced avoid/route detours) when both legs compete for similar liquidity.

3. Inversion Retry Tax (IRT)
   Additional replace/cancel churn triggered by stale local state.

4. Reconciliation Whipsaw Tax (RWT)
   Late truth arrival forces unwind or catch-up in worse microstructure.

Incremental cost:

[ \Delta C_{inv} = DLET + SIT + IRT + RWT ]

3) Data contract (non-negotiable)

Per child lineage, persist:

• parent_id, child_id, replace_chain_id
• t_old_cancel_send_local
• t_old_cancel_ack_local
• t_new_send_local
• t_new_ack_local
• t_first_fill_local, t_last_fill_local
• venue_event_time (if available)
• path_cancel / path_new (gateway+queue identifier)
• qty_old_live_est, qty_new_live_est, qty_overlap
• reason_code (price move, queue-age, urgency, risk)

Without path-level timestamps, inversion cost is easily misclassified as generic market impact.

4) Core metrics

4.1 Intent Inversion Rate (IIR)

Fraction of replace attempts where new ACK lands before old cancel ACK:

[ IIR = P\left(t_{new,ack} < t_{old,cancel_ack}\right) ]

4.2 Dual-Live Duration (DLD)

Per event:

[ DLD_i = \max\left(0,\ t_{old,cancel_ack}^{(i)} - t_{new,ack}^{(i)}\right) ]

Track q50/q90/q95 by venue and symbol-liquidity tier.

4.3 Overlap Notional Ratio (ONR)

[ ONR = \frac{\sum |qty_{overlap} \cdot px|}{\sum |qty_{parent} \cdot px|} ]

4.4 Path Asymmetry Factor (PAF)

[ PAF = \frac{q95(latency_{cancel})}{q95(latency_{new})} ]

A rising PAF is an early warning for inversion regime.

4.5 Reconciliation Whipsaw Cost (RWC)

Cost of post-facto flatten/rebalance after late state convergence.

5) Modeling approach

Use a two-part model:

Layer A — baseline execution cost

(\hat{C}_{base}): spread, local depth, short-horizon vol, queue pressure, participation.

Layer B — inversion residual

[ \hat{\epsilon}_{inv} = f(IIR, DLD, ONR, PAF, RWC, venue, stress) ]

Total forecast:

[ \hat{C}{total} = \hat{C}{base} + \hat{\epsilon}_{inv} ]

Fit quantiles (q50/q90/q95), not mean-only, because inversion damage is tail-heavy.

6) Control state machine

1. SYNCHRONIZED — cancel/new ordering reliable
2. ASYMMETRIC — PAF rising, occasional inversions
3. INVERTED — frequent dual-live windows
4. SAFE_SINGLE_PATH — force conservative sequencing, protect capital

Example transitions:

• SYNCHRONIZED -> ASYMMETRIC if PAF > 1.8 or IIR > 0.08
• ASYMMETRIC -> INVERTED if q95(DLD) and ONR breach thresholds
• INVERTED -> SAFE_SINGLE_PATH if q95 budget burn accelerates or RWC spikes
• De-escalate only after sustained normalization window (hysteresis)

7) Action policy by state

SYNCHRONIZED

• Normal replace cadence
• Standard outstanding-leaf limits

ASYMMETRIC

• Reduce replace frequency
• Increase minimum dwell time before repricing
• Tighten max concurrent replaces

INVERTED

• Prefer amend-in-place (if priority-preserving)
• Suppress non-critical reprices
• Cap overlap notional per symbol

SAFE_SINGLE_PATH

• Enforce cancel-confirm-before-new for selected symbols/venues
• Block urgency escalations driven by unconfirmed residuals
• Route via most deterministic path over fastest path

8) Replay / stress test design

Replay intent-level event streams with controlled perturbations:

• Baseline: original ordering
• Stress A: cancel-path latency +p95 bursts
• Stress B: clustered cancel ACK stalls around volatility spikes
• Stress C: venue-specific cancel throttling, new path unchanged

Evaluate:

• Incremental IS (mean/q90/q95)
• Completion reliability
• Overlap notional incidents
• Reconciliation whipsaw frequency

Promotion rule: no rollout unless q95 tail cost improves (or stays flat with materially lower inversion incidents).

9) Production guardrails

1. Overlap notional ceiling per symbol tier
2. Replace-rate limiter in ASYMMETRIC+
3. Single-path fallback switch for unstable venues
4. Residual trust discount when inversion evidence rises
5. Operator page for sustained INVERTED duration

10) Minimal implementation sketch

if PAF > paf_warn or IIR > iir_warn:
    state = ASYMMETRIC

if q95_DLD > dld_crit or ONR > onr_crit:
    state = INVERTED

if q95_budget_burn > burn_crit or RWC > rwc_crit:
    state = SAFE_SINGLE_PATH

if state == SAFE_SINGLE_PATH:
    enforce_cancel_confirm_before_new()
    cap_overlap_notional()
else:
    apply_replace_rate_limits(state)

cost_forecast = base_cost(features) + inversion_residual(features)

11) Common mistakes

• Treating replace as logically atomic when transport paths are not
• Optimizing only median latency while cancel tails dominate risk
• Ignoring overlap-notional telemetry because gross position looks "eventually correct"
• Letting urgency logic overreact to stale residuals during inversion windows

12) Practical takeaway

When cancel and new paths diverge, execution pays an intent-ordering tax before spread/impact models even start.

If you model inversion explicitly and throttle replace behavior under uncertainty, you usually cut q95 slippage without sacrificing completion stability.