Latency-Distribution-Aware Child-Order Microbatch Slippage Playbook

Date: 2026-03-11
Category: research
Scope: Live execution where wire/ACK jitter is material (equities/futures, venue-agnostic)

1) Why this topic matters

Most schedulers assume each child order is emitted instantly and independently.

In production, that is false:

• orders are often emitted in bursts,
• network + gateway + venue ACK latency has a heavy tail,
• stale decision-to-arrival windows amplify adverse selection.

Result: a "smooth" schedule on paper can become a toxic burst schedule on the tape, increasing slippage tails.

The practical fix is not just "send slower". It is to optimize microbatch size and spacing using the observed latency distribution and alpha half-life.

2) Core idea

At each decision time, choose among emission patterns:

• SINGLE: one child now
• MICROBATCH(k, Δ): k children emitted with spacing Δ (ms)
• HOLD: wait one control step

Use a model that prices both:

1. impact from immediate footprint,
2. stale-arrival risk from latency distribution.

3) Minimal model

Let:

• (q): remaining parent quantity
• (x): candidate child size per slice
• (k): slices in current microbatch
• (\Delta): inter-slice spacing (ms)
• (L): random decision→venue-arrival latency
• (h): signal half-life (ms) for the short-horizon edge
• (\sigma): short-horizon volatility scale
• (D): local displayed+latent depth proxy

3.1 Stale-arrival probability

If edge decays exponentially, stale risk for one slice:

[ p_{stale} = \mathbb{P}(L > h) = 1 - F_L(h) ]

where (F_L) is empirical CDF of latency for this route/venue/time bucket.

3.2 Cost decomposition per slice (buy side)

[ C = C_{spread} + C_{impact} + C_{stale} + C_{underfill} ]

Simple operational parameterization:

[ C_{spread} \approx \frac{1}{2},spread \cdot \mathbb{1}_{aggr} ]

[ C_{impact} \approx a\left(\frac{x}{D}\right)^\beta ]

[ C_{stale} \approx b,\sigma,\sqrt{\mathbb{E}[L]},p_{stale} ]

[ C_{underfill} \approx c,\text{deadline gap penalty} ]

Batch objective (tail-aware):

[ \min_{k,\Delta,x} ; \mathbb{E}[C_{batch}] + \lambda,CVaR_{95}(C_{batch}) ]

subject to participation, risk, and completion constraints.

4) Why microbatching helps (and when it hurts)

Helps when

• latency jitter dominates (p95/p99 far above p50),
• local depth replenishes between slices,
• short edge half-life is longer than inter-slice spacing,
• router can cancel/reprice between slices.

Hurts when

• very urgent completion (deadline cliff),
• volatility shock with collapsing depth,
• reject loops or ACK backlog (engine saturation),
• signal half-life is shorter than actual arrival tail.

Rule of thumb: if (p_{stale}) jumps while depth is thinning, microbatch should shrink to SINGLE or controlled aggression.

5) Feature set for a production model

• latency percentiles by route/venue: p50/p90/p95/p99
• queueing metrics: gateway queue depth, inflight child count
• LOB state: spread, top depth, imbalance, quote age
• toxicity proxy: short-horizon markout / trade-sign pressure
• schedule state: remaining qty, time-to-deadline, participation headroom
• reliability flags: reject rate, cancel-ack lag, sequence-gap flags

Model outputs per candidate action:

• expected slippage (bps)
• q90/q95 slippage
• completion probability by horizon

6) Online controller (practical policy)

Every control step (e.g., 100-300ms):

1. Build candidate actions: SINGLE, MICROBATCH(2..5, Δ∈{20,40,80}ms), HOLD.
2. Score each candidate with expected + tail objective.
3. Apply hard guards:
   • if reject/backlog high → cap k and disable tight Δ,
   • if deadline near → disallow HOLD,
   • if toxicity spike → reduce passive dwell and batch size.
4. Execute best admissible action.

Emergency downgrade path:

• GREEN: full candidate set
• AMBER: smaller k, larger Δ, no HOLD when behind schedule
• RED: single-slice conservative fallback, or bounded aggressive cross

7) Calibration & monitoring

Weekly calibration

• refit latency CDF buckets (symbol-liquidity × venue × session)
• refit impact exponent (\beta) and stale coefficient (b)
• validate quantile calibration (q90/q95 exceedance)

Intraday monitors

• latency drift: p95(L) / baseline
• stale-hit rate: realized fraction where arrival > half-life
• tail health: realized vs forecast q95 slippage
• completion health: deadline miss rate

Auto fallback trigger examples:

• q95 exceedance > 2x target for 20 minutes,
• reject ratio above threshold,
• latency p99 burst beyond risk budget.

8) Backtest / replay design

Use event replay with injected empirical latency traces (not fixed latency):

A/B/C policies:

• A: baseline fixed-slice scheduler
• B: latency-aware size only
• C: full microbatch action policy (size + spacing + hold)

Report by regime:

• mean IS (bps)
• q95/q99 IS (tail)
• completion rate
• cancel/reject load
• operational incidents (fallback triggers)

Promotion criterion: tail reduction with no material completion deterioration.

9) Implementation checklist

• Child-order event schema has full latency timestamps
• Route/venue latency CDF service available in real-time
• Candidate-action scorer returns quantiles, not only means
• Risk guardrails wired (deadline, participation, reject loops)
• Shadow run completed with drift dashboard
• Canary notional limits + rollback switch configured

10) References

• Almgren, R. & Chriss, N. (2000), Optimal Execution of Portfolio Transactions
  https://www.smallake.kr/wp-content/uploads/2016/03/optliq.pdf
• Huang, W., Lehalle, C.-A., Rosenbaum, M. (2015), The Queue-Reactive Model
  https://arxiv.org/abs/1312.0563
• Taranto et al. (2016), Linear models for the impact of order flow on prices I. Propagators
  https://arxiv.org/abs/1602.02735
• Benzaquen, Bouchaud, Donier et al. (survey), Market Impact: Empirical Evidence, Theory and Practice
  https://hal.science/hal-03668669v1/file/Market_Impact_Empirical_Evidence_Theory_and_Practice.pdf

One-line takeaway

If latency is random and heavy-tailed, execution should optimize when and how many child orders to emit per burst—not just how much to trade per minute.