Overnight-Gap + Opening-Auction Slippage Modeling Playbook

Date: 2026-03-08
Category: Research (slippage modeling)
Focus: Modeling execution cost around the open, where overnight information shocks and auction uncertainty dominate microstructure.

1) Why this matters

Open-session execution is not just “normal intraday with wider spreads.” It is a different regime:

• overnight news gets batch-priced,
• latent liquidity shows up discontinuously,
• auction imbalance can flip fast,
• first minutes of continuous trading can re-price again.

If your model ignores this regime break, it underestimates both tail slippage and completion risk.

2) Core decomposition

Treat open execution as a two-stage problem:

1. Auction stage (call auction uncross)
2. Post-open stage (first N minutes continuous trading)

For parent order size (Q):

[ \mathbb{E}[C] = p_{auc},\mathbb{E}[C_{auc}\mid fill] + (1-p_{auc}),\mathbb{E}[C_{post}\mid residual] + C_{opp} ]

Where:

• (p_{auc}): probability of meaningful auction fill,
• (C_{auc}): slippage if auction fills,
• (C_{post}): slippage to clean residual after open,
• (C_{opp}): opportunity cost from delayed/partial completion (alpha decay).

Model q50/q90/q95 for each component, not only mean.

3) Feature contract (pre-open only)

3.1 Overnight shock block

• Gap vs prior close: (G = (P^{ind}{open}-P{prevClose})/P_{prevClose})
• Overnight index-futures move (local + global)
• Overnight realized range / jump proxy
• Corporate/news event flag (earnings, macro release day)

3.2 Auction microstructure block

• Indicative uncross path volatility (std of indicative price updates)
• Indicative imbalance magnitude and sign persistence
• Imbalance sign-flip count (churn proxy)
• Indicative matched volume trajectory (slope + acceleration)
• Auction extension / volatility-interruption flags

3.3 Symbol fragility block

• ADV-relative order size (Q/ADV)
• Typical opening spread percentile
• Pre-open depth proxy (if available)
• Prior 20-day open-to-5min slippage distribution

4) Model architecture

Use a hurdle + conditional quantile stack:

1. Hurdle model: (p_{auc}) (probability residual after auction is below threshold)
2. Auction cost model: (Q_\tau(C_{auc}))
3. Residual post-open model: (Q_\tau(C_{post}))
4. Opportunity-cost model: (Q_\tau(C_{opp})) from expected delay and alpha half-life

Then aggregate into total quantile envelope for control decisions.

Practical tip: Keep auction and post-open models separate. Their errors drift for different reasons.

5) Regime state machine

Define an Open Uncertainty Score (OUS):

[ OUS = w_1|G| + w_2,\sigma(P^{ind}) + w_3,|Imb| + w_4,FlipRate + w_5,Fragility ]

States:

• OPEN-NORMAL: small gap, stable imbalance
• OPEN-NOISY: larger gap or unstable indicative path
• OPEN-STRESS: large gap + imbalance churn + fragility
• OPEN-SAFE: confidence collapse / exchange control event

Control examples:

• OPEN-NORMAL: planned auction participation + standard post-open POV
• OPEN-NOISY: reduce auction clip, reserve residual budget for post-open
• OPEN-STRESS: cap aggression, widen passive patience, stricter tail budget
• OPEN-SAFE: minimum-risk execution only; prioritize survival over schedule

Use hysteresis to avoid flip-flopping during fast pre-open updates.

6) Decision policy (pre-open to +5m)

At T-10m to open:

1. Score (p_{auc}), q90/q95 total cost, and confidence.
2. Choose auction participation ratio (\rho_{auc}).
3. Pre-plan residual schedule for 0–5m (notional ladder + max aggression).

At open print:

1. Recompute residual + updated uncertainty.
2. If realized open dislocation > planned band, switch to stress template.
3. Enforce hard limits: max spread-cross count, max short-horizon participation, max q95 budget burn.

7) Calibration loop

Offline (weekly)

• Refit with recency weighting (open regime drifts quickly)
• Slice diagnostics by gap bucket: small / medium / large
• Validate calibration separately for auction days with extensions/halts

Online (daily)

• q95 exceedance by state (OPEN-NORMAL vs STRESS)
• Auction-fill forecast reliability curve
• Residual-size bias after open (systematic under/over-forecast)
• Budget burn-rate in first 5 minutes

Trigger tighter policy if exceedances cluster in OPEN-NOISY/STRESS.

8) Validation checklist

Before production promotion:

• Better q95 calibration vs single-stage baseline
• Lower first-5m tail losses at same completion target
• No hidden “improvement by underfilling”
• Stable performance on macro-event opens
• Deterministic fallback policy tested in replay

9) Common failure modes

1. Leakage from post-open data into pre-open model -> fake edge
2. Overweighting indicative price levels vs path stability -> fragile forecasts
3. Ignoring exchange control events (auction extensions, VI/halts) -> broken assumptions
4. Cost-only optimization without completion penalty -> operationally unusable
5. No confidence gating -> model acts aggressively when blind

10) Minimal rollout plan

1. Shadow mode (2 weeks): score only, no control actions
2. Advisory mode: recommended (\rho_{auc}) + residual plan, human override
3. Limited auto mode: small notional, strict OPEN-STRESS caps
4. Scale-up: only after stable tail calibration + completion metrics

Always keep a hard SAFE-mode playbook available for event opens.

11) Practical takeaway

The open is a distinct microstructure regime where overnight information, imbalance dynamics, and execution urgency collide. Modeling it as a two-stage uncertainty problem (auction + post-open residual) gives materially better tail control than intraday-style single-stage slippage models.