Multi-Parent Self-Competition Slippage Playbook

Date: 2026-03-02
Category: research (quant execution / slippage modeling)

Why this playbook exists

Many desks model slippage as if each parent order is alone. In production, multiple strategies often trade the same symbol and window:

• alpha rebalance,
• risk hedge,
• intraday signal strategy,
• liquidity transfer/cash target flow.

If these parents execute independently, they compete for the same queue and spread. Result: avoidable slippage inflation even when each strategy looks locally optimal.

This playbook models and controls that hidden tax.

1) Cost decomposition with internal competition

For parent order set (\mathcal{P}_t) on the same symbol/time bucket:

[ C_{total}=C_{market}+C_{self\text{-}competition}-C_{coordination\ benefit} ]

Where:

• (C_{market}): unavoidable market impact + spread + timing cost,
• (C_{self\text{-}competition}): extra cost from our own concurrent flow,
• (C_{coordination\ benefit}): savings from synchronized scheduling/internal crossing.

A practical estimate per parent (i):

[ \Delta C_i \approx \beta_1 \cdot overlap_i + \beta_2 \cdot same_side_pressure_i + \beta_3 \cdot queue_reset_rate_i ]

This turns "we traded too aggressively" into measurable internal externalities.

2) Self-Competition Index (SCI)

Define a live state signal:

[ SCI_t = w_o O_t + w_s S_t + w_q Q_t + w_m M_t ]

• (O_t): parent overlap ratio (same symbol/time),
• (S_t): same-side net pressure from own parents,
• (Q_t): queue reset intensity from own cancel/replace actions,
• (M_t): short-horizon adverse markout after own clustered fills.

Regime ladder:

• GREEN: (SCI < a) (independent execution acceptable)
• AMBER: (a \le SCI < b) (coordination required)
• RED: (SCI \ge b) (centralized scheduler + hard caps)

3) Coordination objective (desk-level)

At each decision cycle, choose participation rates (u_i) for parents (i=1..N):

[ \min_{u} \sum_i \Big(E[C_i(u_i)] + \lambda_{95}Q_{0.95}(C_i) + \lambda_{sla}R_i\Big) +\gamma \sum_{i\neq j} K_{ij}u_i u_j ]

subject to:

• parent deadline/completion constraints,
• risk and compliance limits,
• per-venue throttle limits,
• parent priority floors (must-fill orders).

Interpretation:

• first term: parent-level cost + tail + deadline risk,
• second term: internal cross-impact penalty via (K_{ij}).

If (K_{ij}) is high, those two parents should not hit the tape simultaneously.

4) Execution controls that actually work

A. Staggering policy

When SCI is AMBER/RED, enforce time staggering windows:

• high-priority parent gets first queue opportunity,
• lower-priority parent enters with delayed/low POV schedule,
• avoid same-millisecond same-side sweeps.

B. Netting and internal crossing

Before routing externally:

• net opposite-side parents internally when allowed,
• route only residual net exposure,
• preserve audit trail for internal transfer pricing.

C. Venue partitioning

Assign parents to distinct venue roles under stress:

• parent A: passive lit quoting,
• parent B: midpoint/conditional,
• parent C: residual urgency lane.

This lowers own-order collision probability.

D. Queue-preserving amendment policy

In RED, restrict cancel/replace frequency; prefer amend-keep-priority paths where market rules permit.

5) Minimal production data contract

Per decision event, log:

• parent id, strategy id, symbol, side, residual, deadline,
• concurrent parent set snapshot,
• overlap/same-side pressure/SCI components,
• chosen coordination mode (independent/staggered/netted/centralized),
• child-order actions and venue,
• fill path + short-horizon markouts (1s/5s/30s),
• completion SLA and realized shortfall.

Without parent linkage, self-competition cannot be diagnosed.

6) Validation protocol

Offline counterfactual replay

Compare:

1. baseline independent execution,
2. SCI-aware coordinator.

Required wins (out-of-sample):

• lower mean shortfall,
• lower q95 shortfall,
• no degradation in completion SLA.

Online canary

• start with low-overlap symbols,
• 5% -> 15% -> 30% coordinated flow ramp,
• auto-rollback on q95 or SLA breach.

7) Failure modes and fixes

1. Coordinator improves mean, worsens urgent parent fills
   Fix: hard priority floor + max deferral bound.

2. Too much netting delays true market risk transfer
   Fix: netting TTL and residual force-release rule.

3. Regime flapping (GREEN/AMBER/RED oscillation)
   Fix: hysteresis + minimum hold time per regime.

4. Desk-level gains but strategy-level unfairness complaints
   Fix: transparent transfer-pricing and attribution ledger.

8) 30-day implementation slice

Week 1

• parent-linkable event schema,
• SCI feature computation job.

Week 2

• cross-impact penalty estimator (K_{ij}),
• simulator with overlap perturbation tests.

Week 3

• coordinator service (stagger/net/partition policy),
• shadow output + dashboards.

Week 4

• canary rollout with rollback guards,
• desk-level TCA report: market vs self-competition split.

References to revisit

• Almgren, Thum, Hauptmann, Li (2005), Direct Estimation of Equity Market Impact.
• Kissell (2013), The Science of Algorithmic Trading and Portfolio Management.
• Cartea, Jaimungal, Penalva (2015), Algorithmic and High-Frequency Trading.

TL;DR

A meaningful fraction of slippage can be self-inflicted when multiple parents trade independently. Model that tax explicitly (SCI + cross-parent penalty), coordinate child-order timing/venue/netting, and optimize at desk level rather than strategy silo level.

That is often the cheapest bps you can recover without predicting the market better.