Odd-Lot Reliability Decay at NBBO-Transition Slippage Playbook

Date: 2026-03-11
Category: research (quant execution / slippage modeling)

Why this matters

Modern US equity microstructure exposes a practical trap:

• odd-lot quotes can look like meaningful inside liquidity,
• but their reliability is state-dependent,
• especially around fast NBBO transitions,
• so executable size is often overestimated right before you aggress.

Result: apparent spread quality improves, but realized execution quality deteriorates.

This creates a hidden slippage source I call Transition Mirage Tax (TMT).

Core concept: Reliability-Adjusted Executable Depth (RAED)

Displayed odd-lot depth is not equal to expected fillable depth.

Define:

[ \text{RAED}t = \sum{i \in \text{levels}} q_{i,t} \cdot p^{\text{survive}}{i,t}(\Delta t) \cdot p^{\text{queue-win}}{i,t} ]

where:

• (q_{i,t}): displayed depth (incl. odd-lot contribution),
• (p^{\text{survive}}_{i,t}(\Delta t)): probability quote survives until your order interacts,
• (p^{\text{queue-win}}_{i,t}): probability your child order actually captures queue priority.

Define Transition Mirage Tax:

[ \text{TMT} = \text{RealizedCost} - \text{CostPredictedUsingDisplayedDepth} ]

If TMT is persistently positive during NBBO transition windows, odd-lot reliability is being mispriced.

Where TMT spikes

1. Spread tighten/widen transitions with elevated quote churn.
2. Micro-burst volatility windows around macro headlines or index/futures impulses.
3. Venue-level queue shocks (cancel waves, latency asymmetry, fast repricing).
4. High-priced symbols where odd-lots dominate inside-touch representation.

Detection signals (first dashboard)

1. Odd-Lot Survival Curve (OLSC)
   Survival probability of odd-lot touch quotes by horizon (1ms/5ms/10ms/25ms).

2. Inside-Touch Evaporation Rate (ITER)
   Fraction of odd-lot inside depth canceled/repriced before interaction.

3. Transition Churn Index (TCI)
   Quote updates per millisecond around NBBO state changes.

4. Displayed-vs-Executed Depth Gap (DEDG)
   Predicted fill from displayed depth minus realized fill.

5. Post-Transition Markout Drift (PTMD)
   Short-horizon adverse markout after fills initiated in transition windows.

6. Mirage Contribution Ratio (MCR)
   Share of shortfall attributable to odd-lot reliability misspecification.

If OLSC drops and DEDG rises while TCI is elevated, you are paying TMT.

State machine

1) STABLE_SUPPORT

• OLSC healthy
• TCI low/moderate
• DEDG near baseline

Policy: normal quoting/execution policy with reliability weighting enabled.

2) FRAGILE_ODDLOT

• OLSC decays vs baseline
• ITER rises
• TCI increasing

Policy: down-weight odd-lot depth in sizing model; tighten passive patience horizon.

3) TRANSITION_HAZARD

• frequent NBBO flips
• TCI high and persistent
• DEDG + PTMD worsening

Policy: switch to transition-aware tactic set (smaller clip, stronger recheck, selective crossing only when urgency justifies).

4) SAFE

• projected q95 slippage breach from TMT risk

Policy: deterministic protection mode (hard depth haircut + toxicity guard + tail-budget limiter).

Use hysteresis and minimum dwell-time to avoid over-switching.

Modeling stack

Layer 1 — Quote survival / hazard model

Estimate odd-lot quote lifetime hazard:

[ \lambda_t = h(\text{TCI}, \text{spread state}, \text{micro-vol}, \text{venue}, \text{time-of-day}) ]

Then derive (p^{\text{survive}}(\Delta t)).

Practical choices:

• Cox/GBDT survival,
• piecewise-constant hazard by regime,
• online recalibration by venue-symbol bucket.

Layer 2 — Execution surface with reliability features

Model slippage as:

[ C \sim g(\text{aggression}, \text{RAED}, \text{transition state}, \text{toxicity}, \text{latency}) ]

Track q50/q90/q95 conditioned on transition regime, not just unconditional averages.

Layer 3 — Counterfactual attribution

Replay parent orders with:

• displayed-depth assumption vs RAED assumption,
• transition-aware vs transition-blind policy.

Attribution target: isolate TMT basis points and tail amplification.

Control policy design

1. Reliability haircut on odd-lot depth
   Replace raw depth with RAED in child sizing and participation logic.

2. Transition cooldown gate
   When TCI exceeds threshold, require one additional micro-check before crossing.

3. Clip-size quantization by survivability
   Reduce child size as odd-lot survivability deteriorates.

4. Venue reliability scorecard
   Route weighting reflects venue-specific odd-lot persistence quality.

5. Tail-budget limiter
   If projected q95 TMT exceeds budget, escalate to SAFE mode.

6. Deadline-aware override
   Keep explicit completion floor; avoid over-defensive behavior that causes miss-risk.

Data contract (minimum)

Per decision / order event:

• nanosecond-level decision/send/ack/fill timestamps,
• odd-lot vs round-lot depth snapshot at decision and interaction times,
• NBBO transition flags and transition-window identifiers,
• venue-level quote add/cancel/replace counts,
• queue-position proxy features,
• fill outcome + short-horizon markouts,
• parent linkage for shortfall attribution.

Without transition labeling and odd-lot decomposition, TMT is invisible.

Rollout plan

1. Shadow diagnostics (1-2 weeks)
   Build OLSC/TCI/DEDG dashboard and estimate baseline TMT.

2. Canary (5-10% notional)
   Enable RAED haircuts + transition cooldown gate.

3. Controlled expansion
   Add venue reliability routing weights and tail-budget limiter.

4. Full rollout
   Keep automatic rollback guards.

Rollback triggers:

• completion probability deterioration,
• q95 slippage regression,
• excessive SAFE-mode occupancy.

Common failure modes

• Treating odd-lot inside touch as fully executable.
• Using static haircuts not conditioned on transition state.
• Optimizing average cost while ignoring transition-conditioned tails.
• Missing venue heterogeneity in odd-lot reliability.
• Ignoring interaction between reliability filters and deadline constraints.

Bottom line

Odd-lot visibility improved quote optics, not guaranteed executable liquidity.

To prevent transition-window slippage blowups, execution engines must move from displayed-depth logic to reliability-adjusted depth logic with explicit transition-state control and tail-budget governance.