IOC Reject-Retry Burst Slippage Playbook

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

Why this matters

IOC (Immediate-Or-Cancel) looks “safe” because it avoids queue risk:

• either you fill now,
• or the unfilled remainder is canceled.

But in fast, fragile books, IOC-heavy logic can create a hidden cost loop:

1. send IOC into thin touch,
2. partial/no fill,
3. immediate retry at worse price,
4. repeated urgency escalation,
5. terminal catch-up cost.

This reject-retry burst often behaves like a convex slippage tax, especially near deadlines.

Core concept: IOC Retry Tax (IRT)

Define IRT as the incremental execution cost caused by repeated IOC retries under unstable microstructure.

[ \text{Total Slippage} = \text{Spread/Fee} + \text{Impact} + \text{Delay} + \text{Opportunity} + \text{IRT} ]

IRT is not just “bad luck.” It is a policy-dependent microstructure penalty.

Mechanism map

Typical burst pattern

• IOC #1 sees displayed size that vanishes by arrival.
• Fill ratio is low (or zero), residual stays large.
• Router retries quickly (often with higher aggression).
• Each retry updates reference price in a worsening path.
• Realized cost becomes dominated by retries × drift × widening spread.

Why average models miss it

Mean slippage models often blur together:

• calm single-shot IOC fills, and
• stress-period retry cascades.

You need explicit branch modeling for retry episodes.

Observable metrics

1) IOC Reject Rate (IRR)

[ \text{IRR} = \frac{#\text{IOC attempts with zero fill}}{#\text{IOC attempts}} ]

2) Retry Burst Length (RBL)

Number of IOC attempts in one parent-order micro-episode (e.g., within 1–3 seconds and same residual intent).

3) Retry Gap (RG)

Median inter-attempt time during burst. Very short RG can amplify market-impact signaling and chase risk.

4) Retry Slippage Slope (RSS)

Incremental bps cost per additional retry index within episode.

[ \text{RSS} \approx \frac{\Delta \text{episode cost (bps)}}{\Delta \text{retry index}} ]

5) Terminal Catch-Up Share (TCS)

Fraction of parent-order slippage realized in final completion window after retry bursts.

High TCS means retries deferred cost instead of reducing it.

Modeling approach

Use a two-stage model.

Stage A: Fill/Reject branch model

For each IOC attempt, estimate:

• (P(\text{full fill}))
• (P(\text{partial fill}))
• (P(\text{reject/zero fill}))

with features such as:

• touch depth + depth half-life,
• cancel intensity,
• quote age / feed lag proxy,
• short-horizon volatility,
• attempt index within episode,
• time-to-deadline residual pressure.

Stage B: Episode-level cost model

Condition on episode trajectory:

• branch cost for retries,
• branch cost for fallback aggression,
• missed-opportunity branch when IOC under-completes.

Optimize action policy with tail awareness:

[ \min_a ; \mathbb{E}[\text{Cost}|a] + \lambda,\text{CVaR}_{95}(\text{Cost}|a) + \eta,\text{MissPenalty} ]

where action (a) includes retry timing, size decay, route split, and fallback trigger.

Execution state machine

• STABLE_IOC: low IRR, short bursts, low RSS.
• FRAGILE_TOUCH: IRR and cancel intensity rising; avoid blind rapid retries.
• BURST_RISK: persistent RBL↑ + RSS↑; throttle retry cadence, diversify routes, pre-commit fallback.
• SAFE: deadline pressure + unstable fills; prioritize controlled completion over further retry gambling.

Add hysteresis and minimum dwell time to avoid state flapping.

Practical policy knobs

When state >= FRAGILE_TOUCH:

1. Retry pacing: enforce minimum retry gap (anti-thrash).
2. Size shaping: reduce per-attempt IOC size in fragile depth.
3. Route diversification: avoid repeated collisions on one venue/path.
4. Fallback ladder: predefine IOC -> passive join/improve -> controlled take escalation.
5. Deadline-aware cap: stop low-information retries near cutoff; switch early to deterministic completion mode.

Data contract (minimum)

Per child attempt:

• decision/send/ack/fill/cancel timestamps,
• side/price/size/TIF,
• venue/router path,
• visible depth snapshot at decision and send,
• quote age and spread state,
• markout ladder (100ms/1s/5s/30s),
• parent residual and deadline features,
• attempt index and episode id.

Without event-time quality, retry-episode attribution becomes noisy.

Calibration and monitoring

Weekly

• Refit branch model by liquidity bucket.
• Re-estimate RSS and TCS distribution.
• Re-check calibration for reject probability.

Daily

• Monitor IRR/RBL drift by symbol and time bucket.
• Compare fallback branch cost vs baseline.
• Alert on unusual burst frequency.

Intraday guardrails

• If IRR > threshold and RBL exceeds cap for N windows -> force state >= BURST_RISK.
• If deadline pressure + high RSS -> escalate SAFE and reduce retry budget.

Rollout plan

1. Shadow (2–3 weeks): compute episode metrics and counterfactual policies only.
2. Canary (5–10% flow): enable retry pacing + fallback ladder.
3. Scale by cohort: liquid names first, then thinner names.
4. Promotion gates:
   • q95 slippage non-inferior or improved,
   • completion reliability stable,
   • TCS reduced,
   • no excessive opportunity-loss increase.

Rollback quickly if completion degrades or tail cost widens.

Common failure modes

• Treating every zero-fill IOC as harmless cancellation.
• Retrying too fast with no state awareness.
• Ignoring episode-level tail attribution (focusing only per-attempt averages).
• Delaying deterministic completion until terminal window.

Bottom line

IOC is not automatically low-risk execution.

Under fragile microstructure, repeated IOC retries can become a convex slippage engine. Model the retry episode explicitly, control it with state-aware pacing and fallback rules, and optimize for tail-safe completion, not just first-attempt elegance.

References (starting points)

• Cartea, Jaimungal, Penalva (2015), Algorithmic and High-Frequency Trading.
• Gatheral, Schied, Slynko (2012), transient impact / execution modeling context.
• Bouchaud et al. market microstructure literature on order-book resilience and liquidity fragility.
• Practical venue/routing docs on IOC behavior, reject semantics, and acknowledgment timing.