LULD Band-Proximity & Rejection Slippage Playbook

Date: 2026-03-05
Category: research
Domain: finance / execution / market microstructure / slippage modeling

Why this matters

When a symbol drifts close to its Limit Up/Limit Down (LULD) bands, execution quality changes regime fast:

• displayed depth can look large but become non-executable,
• reject/reprice loops increase,
• queue priority decays while your algo waits,
• reopening auctions can reset the cost baseline.

If your model treats this like normal high-volatility trading, it underestimates tail slippage exactly when control matters most.

1) Cost decomposition for band-stress regimes

Let:

• (P_{arr}): arrival benchmark
• (P_{exec}): realized execution price
• (side\in{+1,-1}): buy/sell sign
• (C): signed slippage in bps

[ C = side\cdot\frac{P_{exec}-P_{arr}}{P_{arr}}\times 10^4 ]

Now split cost into four interpretable components:

[ C = C_{impact}+C_{timing}+C_{reject}+C_{reopen} ]

• (C_{impact}): immediate footprint from aggressing liquidity
• (C_{timing}): drift while waiting / slicing
• (C_{reject}): exchange or venue-level non-execution/repricing overhead near band constraints
• (C_{reopen}): gap and uncertainty cost around pause/reopen auction transitions

In band-proximity regimes, (C_{reject}) and (C_{reopen}) often dominate.

2) Core feature set: distance-to-band is not enough

For each decision time (t):

• Normalized Band Distance (NBD) [ NBD_t=\frac{\min(P_t-LowerBand_t,\ UpperBand_t-P_t)}{P_t} ]
• Band Velocity (BV): signed approach speed toward nearest band
• Band Stickiness (BS): fraction of recent windows spent within (\epsilon) of nearest band
• Reject Intensity (RI): rejects per submitted child order (short window)
• Reprice Churn (RC): cancel/replace rate per second
• Pause Hazard Proxy (PH): model-estimated near-term probability of trading pause

Recommended interaction terms:

• (NBD\times RI): practical executability deterioration
• (BV\times RC): queue-aging acceleration
• (PH\times residual_qty): tail risk on unfinished inventory

3) Model stack

Use a two-layer architecture instead of one monolithic regressor.

Layer A: Regime classifier

Predict state (S_t\in{normal,\ near_band,\ pre_pause,\ reopen}):

[ \Pr(S_t\mid X_t)=g(X_t) ]

Layer B: State-conditional cost model

[ \widehat{C}_t = \sum_s \Pr(S_t=s\mid X_t)\cdot f_s(X_t) ]

Where each (f_s) is trained per regime (quantile model preferred).

Tail-aware objective

[ \min\ \mathbb{E}[C] + \lambda,\mathrm{CVaR}_{95}(C) ]

Near-band execution is mostly a tail-governance problem, not a mean-only optimization problem.

4) Control policy (live execution)

STATE 1 — NORMAL

Trigger: NBD wide, RI stable.

• standard passive/aggressive mix
• normal child-order pacing

STATE 2 — NEAR_BAND

Trigger: NBD below threshold + rising BS.

• reduce passive dwell time
• tighten max retry/reprice count
• increase benchmark refresh cadence

STATE 3 — PRE_PAUSE

Trigger: high PH and accelerating RI/RC.

• prioritize inventory-risk reduction over micro spread capture
• cap passive posting size per venue
• enforce deterministic fallback schedule (no infinite reprice loops)

STATE 4 — REOPEN

Trigger: pause event or auction transition detected.

• switch benchmark: arrival → reopen reference
• widen uncertainty bands for expected fill quality
• use staged re-entry (small canary slices before full participation)

Apply hysteresis and minimum dwell constraints to avoid state flapping.

5) Data contract (must-have)

• Full-depth or robust top-of-book stream with venue timestamps
• LULD band values over time (not only end-of-bar snapshots)
• Child-order lifecycle events: submit/ack/reject/cancel/replace/fill
• Venue and order-type tags for each attempt
• Pause/reopen event markers and auction prints
• Clock sync quality metrics (NTP/PTP drift logs)

Without event-level reject/reprice telemetry, (C_{reject}) cannot be estimated reliably.

6) Validation protocol

Offline replay

• Reconstruct decisions around historical near-band episodes
• Compare baseline vs band-aware policy on:
  • mean slippage
  • p95/p99 slippage
  • reject-loop duration
  • completion ratio by deadline

Shadow mode

• Run policy in parallel without sending real orders
• Track disagreement rate and predicted tail-cost improvement

Canary rollout

• Start with low-notional symbols
• Auto-rollback if any of:
  • CVaR95 worsens beyond threshold
  • reject intensity increases materially
  • deadline completion drops

7) Common failure modes

1. Volatility-only modeling
   High realized vol is not equivalent to band-constrained executability.

2. Ignoring reject loops
   Rejections are not noise; they are state information.

3. No reopen-specific benchmark
   Post-pause costs are misattributed to normal impact/timing.

4. Overfitting rare tails
   Use shrinkage + pooled priors across symbol buckets.

5. Missing telemetry from one venue
   Cross-venue blind spots cause false confidence in control decisions.

8) Minimal implementation checklist

• Add LULD band stream to execution feature store
• Build regime classifier (normal/near-band/pre-pause/reopen)
• Add reject/reprice decomposition to TCA reports
• Train state-conditional quantile slippage models
• Enforce retry budget + deterministic fallback in pre-pause state
• Gate production by CVaR95 and completion-rate SLOs

References to review

• SEC, “Limit Up-Limit Down” Pilot Plan and Associated Events (DERA white paper).
• SEC DERA, “Limit Up-Limit Down” Pilot Plan and Extraordinary Transitory Volatility.
• LULD Plan documentation (NMS Plan participants and amendments): https://www.luldplan.com/
• Zimmermann, K. (SSRN), Price Discovery in European Volatility Interruptions.

If your execution model sees only volatility but not band mechanics, it will fail exactly when the market stops behaving continuously.