Close Auction Imbalance Shock-Absorption Slippage Playbook

Date: 2026-02-27
Category: research (quant execution / slippage modeling)

Why this playbook

The close is where daily risk transfer gets compressed into a short, crowded window. Many desks still treat close execution as:

• “send residual near cutoff,”
• then hope the auction print is fair.

That works on quiet days, but fails exactly when it matters (index rebalances, macro prints, event-driven flow). The result is a recurring pattern:

• under-modeled tail slippage at close,
• residual panic in the last minutes,
• unstable participation between continuous market and auction.

This playbook defines a shock-absorption slippage framework for close execution: estimate imbalance-driven tail risk early, then adapt residual routing before the final minute trap.

Core design principle

Treat close execution as a coupled control problem:

1. Continuous-book leg (T-30m to cutoff): spread/impact/queue dynamics.
2. Auction leg (imbalance publication to uncross): indicative-price drift + fill uncertainty.

Total expected shortfall for a buy metaorder:

[ \widehat{IS}{total} = \widehat{IS}{cont}

• \widehat{IS}_{auction}
• \widehat{IS}_{transition} ]

Where transition captures the handoff error when residual inventory arrives too late into a stressed auction state.

Minimum data contract (must log)

Per symbol per close window (e.g., 1s or imbalance-event buckets):

• decision/sent/ack/fill timestamps,
• child order side/qty/limit/urgency,
• spread, microprice, top-N depth, cancel rate,
• own participation and residual trajectory,
• official imbalance fields (paired qty, imbalance qty/side, indicative uncross price),
• imbalance update cadence and jump size,
• auction eligibility/session flags,
• close print, fill ratio, post-close markouts (5s/30s/5m).

Derived close-specific features:

• Imbalance Pressure Ratio (IPR) = |imbalance qty| / matched qty,
• Indicative Drift Velocity (IDV) = d(indicative price)/dt,
• Uncross Uncertainty Band (UUB) = rolling dispersion of indicative price,
• Residual Compression Ratio (RCR) = residual qty / remaining time.

Model architecture

1) Regime classifier (first)

Classify close state each bucket:

• CALM: low IPR, low IDV, stable updates,
• TILTED: elevated imbalance or drift,
• DISLOCATED: large jumps, unstable indicative path, shrinking time.

Use gradient-boosted trees or lightweight online logistic model with hysteresis. Regime stability matters more than tiny offline AUC gains.

2) Two-head slippage forecaster

Predict two conditional distributions:

Head A — Continuous leg tail cost

[ \widehat{Q}{\tau}(IS{cont} \mid x_t), \quad \tau \in {0.5,0.9,0.95} ]

Inputs: spread/depth/cancel pressure, own POV, residual slope, regime.

Head B — Auction leg tail cost

[ \widehat{Q}{\tau}(IS{auction} \mid z_t) ]

Inputs: IPR, IDV, UUB, imbalance jump frequency, distance to cutoff, residual sent-to-auction.

Use quantile loss + monotonic constraints where sensible (e.g., higher RCR should not reduce tail estimate in stressed regime).

3) Transition penalty model

Explicitly fit transition error:

[ IS_{transition} = f(RCR, \Delta regime, \text{imbalance jump at handoff}) ]

This is the term most teams miss. It explains why “same final print” can still produce materially different realized IS depending on when residual was offloaded.

Online control policy (shock-absorption)

At each step from T-30m to cutoff:

1. Forecast q50/q90/q95 for continuous and auction legs.
2. Compute remaining tail budget:

[ B_{tail}(t) = B_{target} - \widehat{Q}{0.95}(IS{realized+future}) ]

3. Choose residual split between continuous and auction:

[ \min_{u_t, a_t} ; \mathbb{E}[IS] + \lambda \cdot \text{CVaR}_{0.95}(IS) ]

subject to completion and participation caps.

4. Apply regime-dependent actions:

• CALM: harvest passive flow, moderate auction allocation.
• TILTED: increase early residual bleed, reduce end-loaded risk.
• DISLOCATED: prioritize completion safety, cap aggressive sweep size, avoid one-shot last-minute dump.

Practical guardrails

• Residual floor by time: enforce max residual schedule (e.g., by T-5m, residual must be below threshold).
• Imbalance jump breaker: if IPR/IDV spike beyond threshold, freeze aggressive size escalation.
• Auction concentration cap: prevent excessive single-venue/auction dependence on shock days.
• No-flap hysteresis: minimum dwell time before regime downgrade/upgrade.
• Fail-safe fallback: if auction data quality degrades, switch to conservative static close profile.

Validation scorecard

Track weekly by symbol-liquidity bucket:

1. Forecast calibration

   • q50/q90/q95 coverage for continuous + auction heads,
   • transition-penalty residual diagnostics.

2. Execution outcomes

   • close-window IS p50/p95/p99,
   • underfill rate at close,
   • last-2-minute forced aggression frequency.

3. Controller behavior

   • regime occupancy (CALM/TILTED/DISLOCATED),
   • action flip count (anti-churn metric),
   • tail-budget breach incidents.

4. Counterfactual benchmark

   • compare against baseline static-close schedule,
   • report delta on p95 and completion reliability.

Rollout plan

Phase 1 — Shadow (no action)

• Run model and log recommended residual split.
• Compare against live decisions and realized IS.

Phase 2 — Soft control

• Enable only residual-floor and imbalance jump breaker.
• Keep manual override always on.

Phase 3 — Full adaptive split

• Turn on dynamic continuous-vs-auction allocation.
• Canary on liquid names first, then mid-liquidity buckets.

Phase 4 — Governance

• Weekly champion/challenger review,
• Auto-rollback if p95 worsens beyond threshold,
• Monthly event-day replay (rebalance/MSCI/quad-witching analogs).

Failure modes to preempt

• Data-timestamp mismatch between imbalance events and order events

  • Normalize clocks, enforce event ordering checks.

• Overfitting event days

  • Use regime-balanced sampling and strict out-of-time validation.

• Completion obsession causing impact blowup

  • Dual objective (tail risk + completion), not completion-only.

• Model confidence illusion

  • Surface prediction intervals and confidence flags to operator dashboard.

Practical takeaway

Close execution is not “the final minute”; it is a transition system from continuous liquidity into auction uncertainty.

If you explicitly model imbalance shock absorption and transition penalty, you can reduce close p95 slippage without paying chronic underfill penalties.

Suggested references

• Madhavan, A. — Market Microstructure: A Survey.
• Kissell, R. — The Science of Algorithmic Trading and Portfolio Management.
• Bouchaud, J.-P. et al. — empirical works on impact and order flow.
• Exchange-specific close auction rulebooks (for your trading venues) and official imbalance dissemination specs.