Cross-Impact Portfolio Execution Playbook (Production-Oriented)

Date: 2026-02-22
Category: finance / execution research
Purpose: Reduce hidden execution drag when multiple symbols are traded together by modeling cross-impact (trading A moves B) and converting it into practical scheduling/risk controls.

1) Why this matters

Single-name slippage models miss a major live-trading reality: portfolio orders are coupled.

• Selling one large index constituent can pressure close peers/ETF basket.
• Aggressive execution in one symbol can worsen fills in another (same sector, same factor crowding, same liquidity providers).
• During stress, these links strengthen nonlinearly.

Result: per-name models look fine, but portfolio-level implementation shortfall (IS) blows out.

2) Working model (simple enough for production)

Use expected cost for child slices over horizon (t):

[ \mathbb{E}[Cost_t] = \sum_i \alpha_i q_{i,t}

• \sum_i \beta_i q_{i,t}^2
• \sum_{i\neq j} \gamma_{ij} q_{i,t} q_{j,t} ]

• (q_{i,t}): signed participation (or child notional) for asset i at time t
• (\alpha_i): linear components (spread/fees + urgency baseline)
• (\beta_i): self-impact curvature
• (\gamma_{ij}): cross-impact coupling between i and j

Practical interpretation:

• Positive (\gamma_{ij}): same-direction trading in i and j is expensive together.
• Negative (\gamma_{ij}): natural hedge pair; simultaneous flow may be less harmful.

Do not chase theoretical perfection. Stable, slightly biased estimates are better than fragile “exact” ones.

3) Data contract for calibration

For each parent order + child timeline:

1. order_id, symbol, side, parent_notional, start_ts, end_ts
2. child fills: fill_ts, fill_px, fill_qty, venue, passive/aggressive
3. market state snapshots (1s–5s): spread, L1 depth, imbalance, short-term vol
4. benchmark marks: decision price, interval mid, close
5. peer activity proxy: same-sector/factor net signed flow estimate

Derived features:

• normalized signed flow (\tilde q = signed_notional / ADV)
• overlap matrix by time bucket (which names traded concurrently)
• regime labels: normal / stressed (vol-liquidity composite)

4) Estimation approach that survives reality

Step A — fit self-impact first

Per symbol, robust regression/quantile fit for self component (alpha_i, beta_i) by regime.

Step B — estimate sparse cross-impact

Fit (\gamma_{ij}) with strong regularization:

• L1 / elastic-net to force sparsity
• keep only statistically and operationally meaningful edges
• shrink toward sector/factor templates when sample is small

Step C — enforce stability constraints

• Cap (|\gamma_{ij}|) by liquidity sanity bounds
• Ensure cost matrix is near positive-semidefinite (or project to PSD) to avoid optimizer pathologies
• Refit monthly; monitor drift weekly

Production rule: if matrix condition number explodes, fall back to block-diagonal (sector-only cross-impact).

5) Scheduler policy (what to do with model outputs)

At each rebalance step:

1. Compute marginal cost score per symbol:
   • MC_i = dCost/dq_i including cross terms
2. Rank by alpha opportunity / MC_i (alpha-to-impact efficiency)
3. Allocate participation budget with hard caps:
   • per name POV cap
   • sector net-flow cap
   • global execution-risk cap
4. Stagger strongly coupled names:
   • avoid simultaneous aggression on high positive (\gamma_{ij}) pairs
5. Re-evaluate every N minutes or on regime switch

This turns cross-impact from a report artifact into a live control loop.

6) Regime-aware guardrails

Define two live indicators:

• Coupling Pressure Index (CPI): weighted sum of active (\gamma_{ij} q_i q_j)
• Impact Budget Burn (IBB): realized IS / planned IS budget percentile

State machine:

• GREEN: CPI < p70 and IBB < p80
  normal schedule
• AMBER: CPI p70–p90 or IBB p80–p95
  reduce basket concurrency, tighten aggression
• RED: CPI > p90 or IBB > p95
  freeze lowest-alpha legs, execute only critical names defensively

Hysteresis: require 2 consecutive windows improvement before de-escalation.

7) TCA attribution upgrade

Add a cross-impact attribution bucket:

IS_total = self_impact + cross_impact + delay + fees/spread + opportunity

Desk review questions:

• Which symbol pairs generated most cross bucket losses?
• Did scheduler violate “high-coupling stagger” rules?
• Was cross-impact underpredicted in stress regimes?

Without this bucket, teams repeatedly over-blame volatility and under-fix execution policy.

8) Anti-footgun checklist

• Never optimize with unconstrained dense (\gamma) matrix.
• Never reuse normal-regime coefficients during liquidity shock.
• Do not increase urgency simultaneously across correlated legs.
• Track realized vs predicted cross bucket daily.
• Keep a deterministic fallback policy if model confidence drops.

9) 2-week rollout plan

Week 1

• Offline fit self + sparse cross matrix
• Shadow scheduler (no live control), compare realized IS against baseline policy

Week 2

• Enable live in low-risk slice (e.g., 20–30% basket notional)
• Turn on AMBER/RED guardrails + alerts
• Daily post-mortem for coupling hotspots

Success criteria:

• lower p90/p95 portfolio IS (not just mean)
• reduced “bad same-time fills” across coupled names
• stable participation without capacity collapse

10) Key takeaway

Execution edge at portfolio scale is less about perfect single-name models and more about coordinating flow topology. Cross-impact modeling gives the map; regime-aware scheduling and guardrails make it tradable.