Iceberg Reserve-Depletion Nowcasting Slippage Playbook

Date: 2026-03-21
Category: research
Scope: Detecting hidden-liquidity exhaustion before the absorption cliff

Why this matters

Most hidden-liquidity models stop at: “iceberg present vs not present.”

In production, the expensive event is different:

• liquidity looks stable for minutes,
• your schedule leans into that stability,
• hidden reserve suddenly exhausts,
• the next child order pays a sharp impact jump (the absorption cliff).

This playbook focuses on one objective: estimate time-to-exhaustion hazard of hidden reserve at the active price level, and route/schedule before the cliff prints into your implementation shortfall.

Core concept

Treat displayed+hidden liquidity at a level as a finite reservoir with partial observability.

[ D_{eff,t}=D^{disp}_t + R_t - D^{fragile}_t ]

• (D^{disp}_t): displayed queue size
• (R_t): latent reserve (hidden tranche stock)
• (D^{fragile}_t): likely-to-cancel/evaporate component

The control variable is not only expected fill, but:

[ \mathbb{P}(\text{exhaust in }\Delta t \mid \mathcal{F}_t) ]

When this hazard rises, your strategy should switch from “harvest hidden absorption” to “avoid being first fill after exhaustion.”

Observable signals (online, low-latency)

At the best level (or working level) maintain rolling features:

1. Refill Event Count (REC)
   Number of same-price replenishment signatures in lookback window.

2. Inter-Refill Time (IRT)
   Median/quantile of time gaps between refills; stretching IRT often precedes exhaustion.

3. Peak-Size Consistency (PSC)
   Similar displayed refill clip size across events (iceberg-like behavior).

4. Consumed-to-Displayed Ratio (CDR)
   (\text{aggr volume at level} / \text{avg displayed depth}). Persistent high CDR with falling refill intensity is cliff risk.

5. Refill Decay Slope (RDS)
   Local trend of refill intensity (negative slope = reservoir thinning).

6. Queue Survival Residual (QSR)
   Realized queue survival minus model-expected survival. Negative residual indicates hidden support weakening.

7. Pressure Imbalance (PI)
   Recent signed market-order pressure + microprice tilt toward the level.

Latent reserve nowcasting model

Use a compact state-space formulation (fast enough for live use):

State: [ R_{t+1}=\max{0, R_t - C_t + U_t} ]

• (C_t): consumed hidden quantity proxy from executed flow at level
• (U_t): latent refill increment (stochastic, regime-dependent)

Observation model: [ \Pr(\text{refill event}_t=1)=\sigma(\alpha_0 + \alpha^\top x_t + \alpha_R R_t) ]

Exhaustion hazard: [ h_t=\Pr(R_t \le r_{crit} \mid \mathcal{F}_t)=\sigma(\beta_0+\beta^\top z_t) ]

where (z_t) includes REC, IRT, RDS, CDR, QSR, PI, and venue/session covariates.

Practical estimation choices

• Start with penalized logistic hazard (L1/L2) + isotonic calibration.
• If needed, upgrade to particle filter for (R_t) with 50–200 particles.
• Keep model explainable enough for desk + on-call debugging.

Slippage decomposition with cliff-risk term

Extend your shortfall model:

[ IS = Spread + TempImpact + Delay + Fees + \underbrace{CliffRisk}_{\text{new}} ]

Define: [ CliffRisk_t = h_t \cdot \Delta I_t ]

• (h_t): exhaustion hazard now
• (\Delta I_t): expected impact jump conditional on post-exhaust execution

Estimate (\Delta I_t) empirically from episodes where refill pattern breaks and next child crosses through thinner book.

This term prevents over-sizing when hidden-liquidity support is near collapse.

Controller design (state machine)

State A — ABSORBING

Low hazard, strong refill persistence.

• Action: normal passive participation, moderate clip size.

State B — SATURATING

Hazard rising but not critical.

• Action: trim clip size 10–25%, increase inter-child spacing, diversify venues.

State C — CLIFF-RISK

Hazard above critical threshold (e.g., calibrated p80).

• Action: stop leaning on that level; reroute to alternative liquidity sources; reduce passive queue dependence.

State D — EXHAUSTED

Refill signature vanished + impact jump observed.

• Action: immediate protective mode (hard POV cap, tactical pause, aggressive crossing only if urgency/alpha justifies).

Use hysteresis and minimum dwell times to avoid oscillation.

Execution policy knobs tied to hazard

A minimal policy:

[ POV_t = POV_{base}\cdot(1-\gamma_h h_t) ] [ Clip_t = Clip_{base}\cdot(1-\gamma_c h_t) ] [ PassiveBias_t = PassiveBias_{base}-\gamma_p h_t ]

with hard limits:

• (POV_{min} \le POV_t \le POV_{max})
• max child notional
• max expected post-fill markout budget
• kill-switch if hazard + toxicity jointly spike

Calibration workflow (weekly + intraday drift checks)

1. Episode extraction
   • Detect candidate iceberg/refill episodes from level events.
2. Label exhaustion events
   • Refill stop + adverse depth/impact discontinuity within horizon.
3. Fit hazard model
   • Train by venue, session bucket, volatility regime.
4. Calibrate probabilities
   • Reliability diagrams / Brier decomposition.
5. Estimate conditional impact jump (\Delta I)
   • Stratify by symbol liquidity and urgency.
6. Shadow deployment
   • Run controller in observe-only mode before capital exposure.

Monitoring KPIs

Track by symbol bucket and venue:

• p50/p90/p95 implementation shortfall (bps)
• Cliff-loss bps (cost attributable to post-exhaust fills)
• hazard calibration error (ECE/Brier)
• underfill-at-deadline probability
• adverse markout after state C/D transitions
• intervention rate (how often controller switches to cliff protection)

Success criterion: lower p90/p95 shortfall and cliff-loss without unacceptable underfill growth.

Minimal pseudocode

features = featurize(level_events, trades, microprice)
h = hazard_model.predict_proba(features)

if h < th_absorb:
    state = "ABSORBING"
elif h < th_cliff:
    state = "SATURATING"
elif refill_signature_alive(features):
    state = "CLIFF_RISK"
else:
    state = "EXHAUSTED"

params = policy_map[state]
submit_child_orders(params)
log_decision(features, h, state, params)

Anti-footgun guardrails

• Do not infer iceberg from one refill event; require persistence.
• Separate open/close/news regimes; hazard is not portable across regime boundaries.
• Avoid overfitting exact refill clip size (venue microstructure changes).
• Never remove hard risk caps because hazard “looks low.”
• Keep fallback policy deterministic for incident response.

References (starting points)

• Zotikov, D., Antonov, M. (2019), CME Iceberg Order Detection and Prediction (arXiv:1909.09495).
  https://arxiv.org/abs/1909.09495
• Frey, S., Sandås, P. (2009), The impact of iceberg orders in limit order books.
• Moinas, S. (2010), Hidden Limit Orders and Liquidity in Order Driven Markets.
• Bouchaud, J.-P. et al. literature on latent liquidity and impact dynamics.

(Use venue-specific legal/compliance constraints before production rollout.)

Bottom line

Detecting hidden liquidity is useful; detecting when it is about to disappear is where the slippage edge is.

Model reserve-depletion hazard, convert it into explicit control actions, and optimize for tail shortfall—not average fills.