Aging-Aware Passive-Order Timeout Frontier Slippage Playbook

Date: 2026-03-11
Category: research
Scope: Live execution where passive posting is default, but forced completion deadlines exist

1) Why this topic matters

A common production failure is simple:

• passive child order sits too long,
• queue priority decays relative to flow,
• adverse move arrives before fill,
• desk cancels late and crosses aggressively.

This creates a hidden slippage tax: over-wait + panic-cross.

Most routers use fixed TTL (e.g., cancel after N seconds). In reality, optimal timeout is state-dependent and should follow a dynamic frontier, not a single constant.

2) Core framing

For one passive child order, there are three competing outcomes before we decide to cross:

1. Fill (good if markout is healthy)
2. Adverse regime switch (quote fade / micro-trend against us)
3. No event yet (we keep waiting)

Decision variable: timeout horizon (\tau).

• too short (\tau): lose spread-capture opportunities, over-cross
• too long (\tau): higher adverse-selection exposure + completion cliff risk

Goal: choose (\tau) minimizing expected all-in execution cost, including tail risk.

3) Minimal competing-risk model

Let:

• (\lambda_f(t|x)): fill hazard at order age (t), state (x)
• (\lambda_a(t|x)): adverse-event hazard at age (t), state (x)
• (S(t|x)=\exp\left(-\int_0^t (\lambda_f+\lambda_a)du\right)): survival (no event yet)

Then:

[ P_{fill}(\tau|x)=\int_0^{\tau} S(t|x)\lambda_f(t|x)dt ]

[ P_{adv}(\tau|x)=\int_0^{\tau} S(t|x)\lambda_a(t|x)dt ]

[ P_{none}(\tau|x)=S(\tau|x) ]

Expected cost of using timeout (\tau):

[ C(\tau|x)=P_{fill}C_{fill}+P_{adv}C_{adv}+P_{none}C_{cross}(\tau) ]

where:

• (C_{fill}): realized passive fill cost (often spread gain minus drift)
• (C_{adv}): cost if adverse state triggers before fill
• (C_{cross}(\tau)): expected marketable cost for remaining quantity after waiting (\tau)

Control objective (tail-aware):

[ \tau^* = \arg\min_{\tau \in \mathcal{T}} \left(\mathbb{E}[C(\tau|x)] + \eta,CVaR_{95}(C(\tau|x))\right) ]

4) What makes the timeout frontier move

Timeout should shorten when:

• queue-ahead volume is not decaying,
• cancel intensity rises on our side (priority erosion),
• toxicity/markout proxy deteriorates,
• time-to-parent-deadline is small,
• reject/cancel-ack latency is elevated.

Timeout can lengthen when:

• opposite-side depletion probability rises (higher fill chance),
• local depth replenishes behind us,
• short-horizon alpha is favorable and stable,
• completion buffer is large.

Interpretation: timeout is a function (\tau^*(x)), not a global constant.

5) Feature set (production-practical)

State vector (x) examples:

• order age, queue position estimate, queue-ahead shares/contracts
• spread, imbalance, microprice drift, quote-age
• arrival/cancel intensities by side (rolling 100ms–2s windows)
• recent markout proxy (e.g., 200ms/1s signed move)
• child reject ratio, cancel-ack lag, inflight child count
• parent schedule pressure: remaining qty, time-to-deadline, participation gap

A useful split:

• microstructure block: fill/adverse hazards
• schedule block: completion/opportunity cost after timeout

6) Calibration recipe

Offline (daily/weekly)

1. Build child-order lifecycle dataset with right-censoring handled explicitly.
2. Fit competing-risk hazards:
   • baseline: piecewise-constant hazards by age bucket,
   • advanced: Cox / gradient boosting hazard with interactions.
3. Calibrate (C_{cross}(\tau)) from realized aggressive fills conditional on state and urgency.
4. Validate:
   • reliability of (P_{fill}(\tau)),
   • reliability of adverse-event probability,
   • cost-quantile calibration (q90/q95).

Online (intraday)

• apply multiplicative drift correction on hazards via recent residuals,
• re-bucket by venue/session/liquidity regime,
• hard-cap extrapolation in low-sample states.

7) Online controller policy

At each control tick (e.g., 100–300ms):

1. Enumerate candidate (\tau) values (e.g., {150, 300, 500, 800, 1200} ms).
2. Score expected + tail cost for each candidate.
3. Select (\tau^*) with constraints:
   • participation cap,
   • deadline feasibility,
   • risk budget,
   • venue throttles.
4. If current order age (\ge \tau^*), cancel/replace or cross according to urgency tier.

Urgency tiers:

• Green: full frontier optimization (passive favored)
• Amber: reduced max timeout, partial crossing allowed
• Red: timeout collapses to near-zero (completion first)

8) Monitoring and kill-switches

Track continuously:

• realized fill-vs-forecast error by age bucket
• adverse-event hit rate after passive dwell
• timeout-to-cross conversion rate
• q95/q99 implementation shortfall vs forecast
• completion misses near session boundary

Auto fallback triggers:

• model underpredicts adverse hits for >N minutes,
• q95 slippage breach > threshold,
• cancel/replace ACK delays exceed safe limit,
• venue incident flags or abnormal reject bursts.

Fallback mode: fixed conservative timeout + bounded aggression ladder.

9) Backtest / replay experiment design

Use event replay with true order-book event time (not bar aggregates).

Compare:

• A: fixed TTL baseline
• B: static symbol-level TTL
• C: state-aware timeout frontier (this playbook)

Report:

• mean IS (bps)
• q95/q99 IS
• passive fill ratio
• adverse markout after passive fills
• completion rate and deadline misses
• cancel/replace load and reject incidents

Promotion criterion: tail reduction + completion stability, not just mean IS improvement.

10) Implementation checklist

• Child-order lifecycle schema includes fill/adverse/censor outcomes
• Queue-position estimator is versioned and monitored
• Competing-risk model artifacts are reproducible (data snapshot + seed)
• Timeout frontier service has latency SLO and safe defaults
• Intraday drift dashboard wired with alert thresholds
• One-click rollback to fixed-TTL profile available

11) References

• Huang, Lehalle, Rosenbaum (2013/2014), Simulating and analyzing order book data: The queue-reactive model
  https://arxiv.org/abs/1312.0563
• Yu et al. (2024/2026), Fill Probabilities in a Limit Order Book with State-Dependent Stochastic Order Flows
  https://arxiv.org/abs/2403.02572
• Ward et al. (2016/2017), Optimal Execution of Limit and Market Orders with ... Fill Uncertainty
  https://arxiv.org/abs/1604.04963
• Lo, MacKinlay, Zhang (2002), Econometric models of limit-order executions
  https://www.sciencedirect.com/science/article/abs/pii/S0304405X02001344
• Chakrabarty et al. (2006), A Competing Risk Analysis of Executions and Cancellations in a Limit Order Market
  https://doi.org/10.2139/ssrn.934785

One-line takeaway

A passive order should not die at a fixed age; it should expire on a state-aware timeout frontier that balances fill odds, adverse hazard, and completion pressure in real time.