Volatility-Interruption Pre-Trigger Crowding Slippage Playbook

Model the "Auction Cliff" Before the Trigger, Not Just the Auction After It

Why this note: In markets with volatility interruption (VI) style protections, slippage often spikes before the interruption fires. Traders crowd to exit or front-run the potential call-auction transition, spreads gap, queue risk explodes, and your normal continuous-market model underestimates cost.

1) Failure Mode in One Sentence

If your slippage model treats VI as a binary post-trigger event, you miss the high-cost pre-trigger crowding regime where most avoidable damage is already done.

2) Practical Cost Decomposition Near a VI Boundary

For child action (a_t) at time (t):

[ \mathbb{E}[IS_t(a_t)] = C_{spread} + C_{impact} + C_{queue} + C_{transition} + C_{opportunity} ]

Where:

• (C_{spread}): widening + flicker-driven crossing penalty
• (C_{impact}): immediate price concession from urgency
• (C_{queue}): adverse queue aging / priority reset risk
• (C_{transition}): expected cost of switching into auction-mode execution
• (C_{opportunity}): alpha decay if waiting for stabilization

In pre-trigger windows, (C_{queue}) and (C_{transition}) are usually the hidden killers.

3) A Production-Friendly Hazard Formulation

Let (\delta_t) be distance to trigger boundary in bps (or ticks), and (x_t) be microstructure state (imbalance, cancel rate, short-horizon vol, refill speed).

A) Trigger hazard (next (\Delta) ms)

[ \lambda_t = \Pr(\text{VI trigger in }[t,t+\Delta] \mid \delta_t, x_t) ]

Use logistic or survival form:

[ \text{logit}(\lambda_t)=\beta_0+\beta_1\delta_t^{-1}+\beta_2,\text{imbalance}_t+\beta_3,\text{cancelRate}t+\beta_4,\sigma{1s,t} ]

B) Mixture slippage estimator

[ \mathbb{E}[IS_t]= (1-\lambda_t),\mu_{cont}(x_t,a_t) + \lambda_t,\mu_{transition}(x_t,a_t) ]

This alone is often a major calibration upgrade over a single continuous-book model.

4) Regime States You Should Explicitly Encode

• R0 NORMAL: trigger far, standard microstructure
• R1 PRE-TRIGGER WATCH: distance shrinking, spread/refill unstable
• R2 PRE-TRIGGER CROWDING: high cancel bursts, queue churn, urgency clustering
• R3 TRIGGERED TRANSITION: call-auction style handling / constrained routing
• R4 REOPEN STABILIZATION: post-transition noise, temporary depth mirage

Use hysteresis and minimum dwell times so the router does not oscillate.

5) Features That Matter More Than Usual

Boundary proximity

• distance_to_trigger_bps
• distance_velocity_bps_per_sec (approach speed)

Crowd pressure

• cancel_to_trade_ratio_1s
• marketable_flow_share_1s
• same_side_child_burst_rate

Queue fragility

• queue_age_ms
• refill_half_life_ms
• top_levels_depletion_rate

Transition expectation

• historical_pretrigger_to_trigger_prob
• post_trigger_reopen_dispersion_proxy

Missing distance_to_trigger at decision time should block model promotion.

6) Labeling Scheme for Training Data

For each child order, store:

1. Decision-time regime label (R0..R4)
2. Trigger outcome in next windows (e.g., 250ms/1s/5s)
3. Realized IS bps
4. Markouts (1s, 5s, 30s)
5. Fill outcome (fill/cancel/replace/reject)

Then build:

• Mean model (P50) for baseline routing
• Tail model (P90/P97.5) for guardrails
• Hazard model for transition risk

7) Control Policy: Hazard-Aware Execution

NORMAL ((\lambda_t < \tau_1))

• Standard passive/neutral logic

WATCH ((\tau_1 \le \lambda_t < \tau_2))

• Reduce passive timeout
• Narrow acceptable queue-age band
• Cap cancel/replace frequency

CROWDING ((\tau_2 \le \lambda_t < \tau_3))

• Prioritize certainty over marginal spread capture
• Limit child-size to avoid self-impact bursts
• Prefer venues/routes with stable acknowledgments

TRANSITION-RISK ((\lambda_t \ge \tau_3))

• Switch to transition-safe policy profile
• Freeze fragile tactics that depend on stable top-of-book
• Explicitly log exception reason for TCA governance

8) Diagnostics and KPIs

1. Hazard calibration error (predicted vs realized trigger frequency)
2. Pre-trigger tail coverage (P97.5 exceedance in R1/R2)
3. Queue-loss attribution (% IS explained by queue deterioration)
4. Transition surprise rate (trigger without prior WATCH/CROWDING)
5. Late panic share (cost paid in final pre-trigger bucket)

If average IS is flat but pre-trigger tail coverage worsens, your model is silently degrading.

9) Rollout Blueprint

1. Shadow hazard + mixture predictions for 2 weeks
2. Compare against baseline by symbol-liquidity buckets
3. Canary with strict notional caps in R2/R3 only
4. Promote when all hold:
   • Better tail coverage in pre-trigger windows
   • No increase in missed-completion risk
   • Stable reject/retry behavior

10) Common Anti-Patterns

• Treating VI as a post-event-only auction problem
• Using static trigger distance without approach velocity
• Ignoring queue-age/refill dynamics under cancel storms
• Optimizing mean IS while tail risk in R2 explodes
• Training on post-hoc regime labels but routing with weaker real-time labels

11) Fast Implementation Checklist

[ ] Add distance-to-trigger + approach-velocity features to decision telemetry
[ ] Train transition hazard model (short horizon)
[ ] Replace single-cost model with hazard-weighted mixture model
[ ] Add R0..R4 explicit state machine with hysteresis
[ ] Attach tail-coverage and transition-surprise alerts to deployment gates
[ ] Run canary with notional and tactic constraints in high-hazard states

References

• Cartea, Á., Jaimungal, S., Penalva, J. (2015), Algorithmic and High-Frequency Trading.
• Gould, M. D., et al. (2013), Limit Order Books (Quantitative Finance review).
• Lipton, A., Pesavento, U., Sotiropoulos, M. (2013), Trade arrival dynamics and quote imbalance in a limit order book.
• Bacry, E., Mastromatteo, I., Muzy, J.-F. (2015), Hawkes processes in finance.

TL;DR

Model pre-trigger crowding as a distinct hazard-driven regime. A transition-aware mixture model (continuous vs transition cost) plus state-machine controls usually reduces tail slippage more than tuning one global impact curve ever will.