Phantom Liquidity & Cancel-Intensity-Adjusted Slippage Playbook

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

Why this matters

Displayed depth is not guaranteed depth.

In stressed or highly competitive symbols, top-of-book size can vanish exactly when you route or join. If your slippage model treats posted depth as executable depth, it underestimates adverse selection and queue-loss risk.

This playbook models phantom liquidity: visible size that has low conditional survival when you need it.

1) Core decomposition

Let:

• (P_{arr}): arrival benchmark price
• (P_{exec}): realized execution price
• (side\in{+1,-1}): buy/sell sign
• (D_{vis}(t)): visible depth at touch at decision time
• (D_{eff}(t)): expected executable depth after cancellation hazard

Observed slippage (bps):

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

Define Phantom Ratio:

[ \phi(t)=1-\frac{D_{eff}(t)}{D_{vis}(t)} ]

• (\phi\approx 0): displayed depth is reliable
• (\phi\to 1): displayed depth is mostly phantom

Then model expected cost as:

[ \mathbb{E}[C\mid X]=g\big(X_{base},\phi,\lambda_{cancel},\lambda_{trade},\text{burst},\text{markout}\big) ]

2) Estimating executable depth

Use short-horizon competing hazards at each level:

• cancellation hazard (h_c)
• trade-through/consumption hazard (h_t)
• queue-position advancement hazard (h_q)

For horizon (\Delta):

[ S(\Delta)=\exp\Big(-\int_0^{\Delta}(h_c(u)+h_t(u))du\Big) ]

Approximate effective depth:

[ D_{eff}(\Delta)\approx D_{vis}\cdot S(\Delta)\cdot \Pr(\text{you reach fill rank by }\Delta) ]

Key point: two books with same (D_{vis}) can have radically different (D_{eff}).

3) Feature set for phantom-liquidity states

Microstructure reliability features

• Cancel-to-Trade Ratio (CTR) over 100ms/1s windows
• Depth Half-Life (DHL): median time for posted touch depth to decay by 50%
• Replenishment Delay (RD): time to refill after touch depletion
• Order-Book Flicker (OBF): quote updates per second at top 2 levels
• Venue Fade Skew (VFS): cancellation asymmetry by venue and side

Toxicity / outcome features

• immediate markout ladder (100ms, 1s, 5s)
• signed trade imbalance
• microprice drift vs quoted midpoint

Infrastructure features

• ack/cancel round-trip latency percentile
• throttle/reject counts
• sequencing-gap indicators

4) State machine for execution control

STATE A — STABLE_DEPTH

Trigger: low CTR, long DHL, low OBF.

• trust passive queueing more
• normal child slice sizing

STATE B — FRAGILE_DEPTH

Trigger: rising CTR + shortening DHL.

• reduce passive wait timeout
• increase quote revalidation cadence
• tighten max child notional per wave

STATE C — PHANTOM_DEPTH

Trigger: high (\phi), frequent fade-before-fill events, worsening markouts.

• downweight displayed depth in routing objective
• shift mix toward opportunistic taker clips
• enforce stricter adverse-selection budget

STATE D — DISLOCATED

Trigger: extreme OBF + unstable replenishment + reject/throttle pressure.

• safe-mode participation caps
• deterministic fallback schedule
• optional manual supervision for large residuals

Use hysteresis + minimum dwell times to avoid rapid flapping.

5) Model objective

Optimize mean and tail jointly:

[ \min ; \mathbb{E}[C] + \alpha,\mathrm{CVaR}_{95}(C) + \beta,\Pr(C>c^*) ]

Where (c^*) is your per-order slippage risk budget.

This prevents “good average bps” while phantom-depth regimes blow up p95.

6) Data contract (minimum viable)

• full depth snapshots + incremental updates (top 5–10 levels minimum)
• order lifecycle events (submit/ack/replace/cancel/fill)
• venue-tagged fill prints and queue-position proxies
• synchronized sub-ms timestamps across market + execution feeds
• markout ladder by child order

Without reliable event alignment, phantom estimates become noise.

7) Validation protocol

Offline replay

• reconstruct decisions with naive (D_{vis})-only vs (D_{eff})-aware routing
• compare mean/p95/CVaR slippage and completion reliability
• segment by symbol liquidity tier and volatility regime

Shadow mode

• score phantom state in real time without order-policy change
• monitor disagreement between old/new policies

Canary rollout

• start with low notional bucket
• rollback if any guardrail breaches:
  • CVaR95 deterioration
  • completion ratio drop
  • reject/throttle escalation

8) Failure modes to avoid

1. Depth illusion lock-in
   Treating posted size as guaranteed liquidity.

2. No cancel-hazard conditioning
   Same depth value used across calm and stressed regimes.

3. Tail-blind optimization
   Mean improves, tail risk explodes.

4. Cross-symbol pooling abuse
   Phantom behavior is highly symbol/venue specific.

5. Latency ignored in inference
   By the time your decision lands, the inferred book is stale.

9) Minimal implementation checklist

• Compute CTR, DHL, OBF, RD, and per-venue fade metrics daily
• Build (D_{eff}) estimator from competing hazards
• Train slippage model with phantom features + markout labels
• Add execution state machine with hysteresis
• Enforce canary rollback gates on p95/CVaR/reliability
• Publish dashboard: displayed depth vs effective depth gap

References to review

• Bouchaud, J.-P., Farmer, J. D., & Lillo, F. (2009), How Markets Slowly Digest Changes in Supply and Demand.
• Cartea, Á., Jaimungal, S., & Penalva, J. (2015), Algorithmic and High-Frequency Trading.
• Abergel, F., Anane, M., Chakraborti, A., Jedidi, A., & Toke, I. M. (2016), Limit Order Books.
• Moallemi, C. C., & Sağlam, M. (2013), The Cost of Latency in High-Frequency Trading.
• Gould, M. D., Porter, M. A., Williams, S., McDonald, M., Fenn, D. J., & Howison, S. D. (2013), Limit Order Books (Quantitative Finance survey).

If you don’t discount for cancellation intensity, your model prices liquidity that disappears on contact.