Dark-Pool Midpoint Toxicity & Markout Slippage Playbook

Date: 2026-03-05
Category: research
Domain: finance / execution / market microstructure / dark liquidity

Why this matters

Midpoint dark execution is often treated as “free spread improvement.”

That is incomplete.

A midpoint fill can still be expensive when:

• the reference quote is stale,
• contra flow is informed,
• or your order is repeatedly pinged before adverse moves.

So the right question is not “did I get midpoint?” but:

“Did midpoint save more than post-trade markout cost?”

Core cost decomposition

For a buy child order (sell is symmetric), define effective dark cost in bps:

[ C_{dark} = C_{arrival\to fill} + C_{markout} + C_{opportunity} - S_{midpoint} ]

Where:

• (S_{midpoint}): half-spread savings vs crossing lit touch,
• (C_{markout}): adverse price move after fill (signed markout),
• (C_{arrival\to fill}): drift while waiting for dark fill,
• (C_{opportunity}): residual urgency cost if fill probability is low.

Dark routing is only beneficial when expected savings exceed toxicity + waiting costs.

1) Markout ladder as primary toxicity lens

Track signed markout at multiple horizons:

[ M_{\tau} = side \cdot \frac{P_{\tau}-P_{fill}}{P_{fill}} \times 10^4 ]

• (side=+1) for buys, (-1) for sells,
• (\tau) ladder: 10ms, 100ms, 1s, 5s, 30s (venue-dependent).

Interpretation:

• positive (M_{\tau}): favorable post-fill drift,
• negative (M_{\tau}): adverse selection.

Use both median and tail (p10/p5) markouts; means alone hide toxic bursts.

2) Midpoint Toxicity Score (MTS)

Build a compact routing feature:

[ MTS = a_1\tilde{M}{100ms} + a_2\tilde{M}{1s} + a_3\tilde{M}{5s} + a_4\tilde{Q}{stale} + a_5\tilde{J}_{micro} ]

Suggested normalized components:

• (\tilde{M}_{\tau}): z-scored signed markouts by symbol/time bucket,
• (\tilde{Q}_{stale}): stale-reference proxy (quote age / update lag / crossed-feed disagreement),
• (\tilde{J}_{micro}): micro-jump risk proxy (short-horizon realized variance, imbalance shocks).

Calibrate by:

• symbol liquidity bucket,
• session phase (open/mid/close),
• volatility regime.

Do not pool all names into one global MTS model.

3) Dark-vs-lit expected cost switch

At each decision point, compare:

[ \mathbb{E}[C|dark] = p_f\cdot C_{dark,fill} + (1-p_f)\cdot C_{dark,miss} ]

[ \mathbb{E}[C|lit] = C_{lit,now} ]

Route dark only when:

[ \mathbb{E}[C|dark] + \delta < \mathbb{E}[C|lit] ]

• (p_f): near-term dark fill probability,
• (\delta): safety margin for model error / stale-state risk.

This avoids the common anti-pattern: “always try dark first” regardless of toxicity regime.

4) Execution state machine

STATE A — DARK_FAVORABLE

Trigger: MTS low toxicity + stable fill probability.

• prioritize midpoint dark slices,
• keep lit fallback with moderate timeout,
• monitor 100ms/1s markout drift.

STATE B — DARK_GUARDED

Trigger: mild toxicity deterioration or weaker fill odds.

• reduce dark child size,
• shorten dark timeout,
• pre-arm lit continuation path.

STATE C — DARK_TOXIC

Trigger: sustained negative near-touch markouts / stale quote indicators.

• sharply reduce dark exposure,
• shift to lit/taker or protected displayed liquidity,
• tighten max wait budget.

STATE D — DISLOCATED

Trigger: severe quote-staleness bursts, jumpy microstructure, unstable fill quality.

• enter safe mode: minimal dark probing,
• use deterministic fallback schedule,
• require human review for large residuals.

Use hysteresis to prevent state flapping.

5) Data contract (minimum viable)

• fill-level timestamps with sub-millisecond consistency where possible,
• reference NBBO/mid snapshots around fill event,
• quote-age and feed-sequence integrity markers,
• venue/strategy tags (midpoint peg type, conditional logic, min quantity),
• child-order lifecycle (submit/ack/fill/cancel/timeout),
• post-fill markouts by horizon ladder,
• urgency context (residual size, deadline, alpha decay proxy).

If clock sync is weak, toxicity attribution is mostly noise.

6) Validation protocol

Offline

• segment by regime (volatility, spread, quote-age quality),
• evaluate p50/p90/p95 shortfall and p10 markout,
• include missed-fill opportunity cost, not just filled dark trades.

Shadow

• run route decision model in parallel without changing live routing,
• compare model-implied route with realized outcomes,
• monitor false-toxic and false-safe rates.

Canary

• low notional + tight guardrails,
• automatic rollback triggers:
  • p95 shortfall degradation,
  • near-touch markout tail worsening,
  • excessive residual carry from dark timeouts.

7) Frequent failure modes

1. Midpoint fill-rate worship
   High fill rate can coexist with poor markouts.

2. Single-horizon toxicity metric
   100ms may look fine while 1s/5s are strongly adverse.

3. Ignoring quote staleness
   Midpoint quality depends on reference freshness.

4. No opportunity-cost accounting
   Dark misses are not free when alpha decays.

5. No regime segmentation
   One threshold for all symbols/sessions fails in production.

8) Minimal implementation checklist

• Build markout ladder (10ms→30s) for every dark fill
• Add MTS feature pipeline with quote-staleness inputs
• Estimate dark fill probability and missed-fill cost jointly
• Deploy dark-vs-lit expected-cost switch with safety margin (\delta)
• Add state machine + hysteresis + rollback triggers
• Monitor p95 shortfall + p10 markout + timeout carry together

References to review

• Zhu, H. (2014), Do Dark Pools Harm Price Discovery? (Review of Financial Studies).
• Shkilko, A. & Sokolov, K. (2023), Sharks in the Dark: Quantifying HFT Dark Pool Latency Arbitrage (Journal of Economic Behavior & Organization).
• Venue-level midpoint / markout methodology notes (e.g., near-touch markout conventions) for implementation details.

Midpoint is a price level, not a quality guarantee. In dark routing, markout-aware regime control is the difference between hidden alpha and hidden tax.