Index-Futures Spillover Slippage Modeling Playbook

Date: 2026-03-08
Category: Research (slippage modeling)
Focus: Modeling and controlling execution slippage in single names when index-futures flow and basis dislocations spill into cash microstructure.

1) Why this matters

Many equity execution models assume each symbol’s short-horizon slippage is mostly local (spread, depth, queue, own participation). In practice, that fails during:

• index futures bursts (open, macro timestamps, close ramps),
• basis dislocations,
• ETF create/redeem pressure,
• cross-venue risk-off de-leveraging.

In these windows, cash-book conditions are partly downstream of derivatives flow, so local-only slippage models underprice tails and respond too late.

2) Core idea

Model expected child-order cost as:

[ \text{Slippage}{t,h} = f{\text{local}}(X_t) + f_{\text{spill}}(Z_t) + f_{\text{interaction}}(X_t, Z_t) + \epsilon ]

Where:

• (X_t): local microstructure features (spread, depth slope, queue pressure, cancel intensity, own POV)
• (Z_t): derivatives/portfolio spillover features (futures return shock, futures OFI, basis z-score, ETF flow proxy, cross-asset realized vol)
• (f_{\text{interaction}}): nonlinearity (e.g., futures shock * thin local depth)

Use a quantile-first target (q50/q90/q95) instead of mean-only, since spillover mostly explodes upper tails.

3) Data contract

3.1 Required event clocks

• Exchange timestamp (if available)
• Gateway receive timestamp
• Strategy decision timestamp
• Child order sent/ack/fill timestamps

Maintain strict time quality flags; stale or uncertain clocks should lower model confidence.

3.2 Join domains

1. Cash symbol microstructure (L1/L2, queue stats, cancels/trades)
2. Index futures stream (best bid/ask, trades, order-flow imbalance)
3. ETF / basket pressure proxies (premium-discount, implied rebalance pressure)
4. Volatility and regime context (intraday RV, event windows)

Resample to a common short horizon (e.g., 100ms / 250ms / 1s) with explicit missingness markers.

4) Feature set

4.1 Local (X)

• Relative spread (bps)
• Effective depth at top k levels
• Queue imbalance / microprice skew
• Cancel-to-trade ratio and refill half-life
• Own participation and recent aggressive volume

4.2 Spillover (Z)

• Futures return burst: (r^{fut}_{t,\Delta})
• Futures OFI shock percentile
• Basis dislocation z-score: [ B_t = \frac{P^{cash_index}_t - \beta P^{fut}_t}{\hat\sigma_B} ]
• Cross-impact proxy: weighted peer move excluding target symbol
• ETF pressure proxy: premium/discount + primary-flow indicators (when available)

4.3 Interaction terms

• futures_shock × low_depth
• basis_z × cancel_intensity
• futures_OFI × own_POV

These terms usually explain why similar local books produce very different realized costs.

5) Modeling stack

Use a two-layer stack:

1. Baseline local model: robust linear/GBDT quantile model on (X)
2. Spillover residual model: predict residuals with (Z) and ((X,Z)) interactions

Final quantile estimate: [ \hat Q_{\tau} = \hat Q^{local}{\tau} + \hat Q^{spill}{\tau} ]

Practical note: this decomposition helps diagnose whether slippage miss came from local-book misfit vs cross-asset shock.

6) Regime state machine (for execution controls)

Define a Spillover Stress Score (SSS) combining normalized futures shock, basis z, and local fragility:

[ SSS_t = w_1|r^{fut}_{burst}| + w_2|B_t| + w_3\text{Fragility}_t ]

States:

• NORMAL: SSS < a
• WATCH: a ≤ SSS < b
• STRESS: b ≤ SSS < c
• SAFE: SSS ≥ c or model confidence collapse

Control policy example:

• NORMAL: standard adaptive POV
• WATCH: reduce clip size, widen passive patience bands
• STRESS: cap aggression, throttle participation, avoid fragile venues
• SAFE: freeze high-risk tactics, execute only minimal risk-reducing flow

Use hysteresis (enter at high threshold, exit at lower threshold) to prevent state flapping.

7) Calibration and monitoring

7.1 Offline (weekly)

• Refit feature transforms and model parameters
• Re-evaluate feature importance drift
• Stress-test on top 5% futures shock windows

7.2 Online (intraday)

• Quantile coverage tracking (target vs realized exceedance)
• Tail breach rate in STRESS/NORMAL separately
• Spillover attribution ratio: [ SAR = \frac{|\hat Q^{spill}{0.95}|}{|\hat Q{0.95}| + \epsilon} ]

If SAR spikes and coverage worsens, tighten controls or degrade to safer policy.

8) Validation checklist

Before promotion:

• q95 calibration error improves vs local-only baseline
• Tail exceedances reduced in high futures-shock windows
• Completion risk does not blow out beyond desk limits
• State machine transitions are stable (low flap rate)
• Rollback and SAFE-mode triggers tested in replay

9) Common failure modes

1. Clock misalignment between futures and cash streams -> false lead-lag signals
2. Feature leakage from post-fill windows -> fake model gains
3. Overfitting event days -> poor normal-day behavior
4. No completion constraint -> slippage improves by simply not filling
5. Control overreaction -> aggressive throttling causes opportunity-cost blowups

Guardrails: time-quality checks, strict as-of joins, dual objective (cost + completion), and canary rollout with hard rollback rules.

10) Minimal rollout plan

1. Shadow mode (1–2 weeks): score only, no control actions
2. Assist mode: operator-visible recommendations, manual acceptance
3. Limited auto mode (small notional): WATCH/STRESS controls enabled
4. Scaled mode: expand notional only after stable q95 and completion metrics

Always keep a deterministic fallback policy available for SAFE state.

11) Practical takeaway

When index-futures flow is driving local books, local-only slippage models become structurally blind. A spillover-aware stack (with explicit regime controls and tail calibration) converts that blind spot into measurable, controllable execution risk.