Basket Cross-Impact No-Arbitrage Slippage Model Playbook

Date: 2026-03-11
Category: knowledge
Scope: Multi-name basket execution where single-name slippage models underperform due to cross-impact

1) Why this matters in production

Single-name impact models often fail on basket days (rebalance, sector rotations, stat-arb de-risking) because they ignore cross-impact:

• buying one stock can push correlated names,
• your hedge leg can pollute alpha-leg execution,
• child-order sequencing across names changes total cost materially.

If you only model diagonal (self-impact), scheduler decisions are systematically biased in crowded regimes.

2) Core objective

Build a basket slippage model that is:

1. cross-aware (captures off-diagonal impact),
2. stable (does not explode from noisy covariance estimates),
3. no-arbitrage-consistent (avoids impact structures implying manipulation loops),
4. operational (can run intraday with guardrails).

3) Minimal model structure

Let basket of (N) names have signed participation vector (u_t \in \mathbb{R}^N), with positive for buy pressure and negative for sell pressure.

Mid move decomposition:

[ \Delta m_t = \underbrace{B_t x_t}{\text{exogenous move}} + \underbrace{\sum{\tau\le t} G_{t-\tau} K u_\tau}_{\text{impact (self + cross)}} + \epsilon_t ]

• (x_t): exogenous features (market/sector/factor shocks),
• (K): instantaneous impact matrix (diagonal=self, off-diagonal=cross),
• (G_h): decay kernel at lag (h) (scalar or low-rank matrix form),
• (\epsilon_t): residual microstructure noise.

Parent-order expected cost (quadratic approximation):

[ \mathbb{E}[C] \approx \sum_t u_t^\top \Lambda_t u_t + \lambda \cdot \mathrm{Risk}(u_{1:T}) + \eta \cdot \mathrm{CompletionPenalty} ]

where (\Lambda_t) is derived from (K) and decay assumptions.

4) No-arbitrage / anti-manipulation constraints

In practice, unconstrained cross-impact fits are noisy and can imply impossible round-trip profits. Use constraints before promotion:

1. Symmetric core: use (K_s=(K+K^\top)/2) as execution core.
2. PSD projection: project (K_s) onto positive-semidefinite cone (eigenvalue clipping).
3. Diagonal floor: enforce minimum self-impact per name (avoid zero-impact artifacts).
4. Cross bound: (|K_{ij}| \le \rho\sqrt{K_{ii}K_{jj}}), (\rho\in[0,1)).
5. Decay monotonicity: fitted (G_h) non-increasing and non-negative.

These constraints are blunt but robust for live systems.

5) Calibration pipeline (practical)

5.1 Data unit

Use synchronized event buckets (e.g., 250ms / 1s):

• child-order signed qty by name/venue,
• prevailing spread, depth, queue proxies,
• factor returns (market/sector/style),
• realized markouts (1s, 5s, 30s, 5m).

5.2 Two-stage estimation

1. De-factor first: regress returns on factor shocks, keep residual move.
2. Impact fit second: regress residual move on lagged signed flow matrix.

This prevents over-attributing macro/factor motion to own impact.

5.3 Regularization

• Elastic-net or ridge on off-diagonals,
• shrink toward block structure (sector/factor clusters),
• enforce sparse tail (ignore tiny unstable links).

5.4 Intraday refresh

Keep structure fixed intraday, refresh only scale multipliers by regime:

• spread regime,
• realized vol bucket,
• crowding proxy (same-side market flow pressure).

6) Scheduler integration

Given remaining inventory vector (q_t), optimize next child schedule with cross-aware marginal cost:

[ \mathrm{MC}i(t)=\partial C/\partial u{i,t} ]

Policy implications:

• If two names have strong positive cross-impact and same-side inventory, avoid simultaneous aggression spikes.
• If hedge leg cross-hurts alpha leg, shift hedge timing to lower overlap windows.
• Use sector staggering for large baskets to reduce endogenous crowding waves.

A small sequencing change often beats raw aggression increase.

7) Online guardrails (must-have)

1. Condition-number cap: if impact matrix ill-conditioned, fallback to block-diagonal.
2. Cross-share cap: bound fraction of predicted cost from off-diagonal terms.
3. Regime-break detector: if markout residual z-score > threshold, freeze cross term.
4. Kill switch: revert to diagonal-only model during venue outages/data lag.
5. Canary routing: enable cross-aware policy on subset of symbols before full rollout.

8) TCA scorecard for promotion

Report both aggregate and decomposed cost:

• total IS (bps),
• self-impact component,
• cross-impact component,
• residual/exogenous drift,
• completion rate and deadline miss,
• q95/q99 tail slippage.

Promotion criterion (example):

• cross-aware policy improves q95 IS by >= 8%,
• no worsening of completion SLA,
• residual drift attribution remains stable across regimes.

9) Common failure modes

1. False cross-links from asynchronous timestamps
   Fix: tighter clock alignment + lead/lag diagnostics.

2. Factor contamination mistaken as cross-impact
   Fix: stronger de-factor stage and event controls.

3. Overfitted off-diagonals
   Fix: stronger shrinkage + PSD projection + turnover penalty.

4. Regime portability failure
   Fix: regime-conditioned scaling, not full intraday refit.

5. Control instability from noisy marginal costs
   Fix: hysteresis bands and max delta-aggression per step.

10) Implementation checklist

• Unified basket event schema (synchronized timestamps)
• Factor residualization layer in feature pipeline
• Cross-impact estimator with PSD projection utility
• Regime scaler service (vol/spread/crowding)
• Scheduler objective includes off-diagonal terms
• Real-time health panel (condition number, residual drift, cross-share)
• Hard fallback path to diagonal model
• A/B framework with parent-order randomization and rollback gates

11) References

1. Benzaquen, M., Mastromatteo, I., Eisler, Z., & Bouchaud, J.-P. (2018). Cross-impact and no-dynamic-arbitrage. Quantitative Finance.
   https://arxiv.org/abs/1612.07742

2. Mastromatteo, I., Benzaquen, M., Eisler, Z., & Bouchaud, J.-P. (2017). Trading Lightly: Cross-Impact and Optimal Portfolio Execution.
   https://arxiv.org/abs/1702.03838

3. Gatheral, J. (2010). No-Dynamic-Arbitrage and Market Impact.
   https://papers.ssrn.com/sol3/papers.cfm?abstract_id=1292353

4. Almgren, R., & Chriss, N. (2000). Optimal Execution of Portfolio Transactions.
   https://www.smallake.kr/wp-content/uploads/2016/03/optliq.pdf

One-line takeaway

For basket trading, treating slippage as purely single-name is a structural miss—cross-impact with no-arbitrage constraints gives better tail-cost control without sacrificing completion reliability.