Microprice Divergence Exhaustion Slippage Playbook

Date: 2026-03-10
Category: research
Scope: Intraday execution for liquid equities/futures where microprice pressure often overshoots before reverting.

Why this matters

Many execution engines treat one-sided microprice pressure as a simple “avoid passive, cross now” signal. In production, that often overpays:

• Early in pressure build-up, crossing can be correct.
• Late in pressure build-up, crossing often buys the local peak of toxicity.
• A short exhaustion window frequently appears where queue risk is still real, but immediate adverse selection decelerates.

This note turns that intuition into a measurable slippage control loop.

Core hypothesis

Slippage is not only a function of spread/size/volatility. It is strongly conditioned by where we are in the pressure lifecycle:

1. Pressure Build: toxicity accelerating (don’t wait too long).
2. Exhaustion Window: toxicity still high but incremental damage decelerating.
3. Reversion/Refill: crossing urgency can be reduced; passive/improve regain edge.

Goal: avoid paying the “last expensive ticks” of panic-taking while preserving completion reliability.

Signal stack

Use event-time features (e.g., 100–500ms bars for fast symbols, 1–3s for slower symbols).

1) Normalized Microprice Divergence (NMD)

[ \text{NMD}_t = \frac{\text{MicroPrice}_t - \text{Mid}_t}{\max(\epsilon, \text{Spread}_t)} ]

Interpretation:

• |NMD| near 0: balanced book.
• Large |NMD|: strong imbalance pressure.

2) Pressure Acceleration (PA)

[ \text{PA}_t = \Delta \text{NMD}_t ]

Positive and rising PA means pressure is still building (high chase risk if you wait).

3) Toxicity Deceleration Ratio (TDR)

Track short-horizon markout slope vs prior slope:

[ \text{TDR}t = \frac{|\nabla \text{Markout}{0\to2s}|}{|\nabla \text{Markout}_{-2s\to0}| + \epsilon} ]

• TDR > 1: toxicity accelerating.
• TDR < 1: toxicity decelerating (candidate exhaustion window).

4) Refill Elasticity Index (REI)

After top-level depth depletion, estimate refill speed:

[ \text{REI}t = \frac{\text{RefilledDepth}{\tau}}{\text{SweptDepth}_{\tau} + \epsilon} ]

Higher REI suggests impact recovery is improving.

5) Sweep Burst Intensity (SBI)

Short-window count/size of aggressive market sweeps. Helps distinguish structural move vs transient burst.

Regime state machine

State A — BALANCED

• |NMD| low, PA near zero.
• Baseline schedule policy.

State B — PRESSURE_BUILD

• |NMD| high, PA positive, TDR >= 1.
• Increase urgency (but cap max aggression).

State C — EXHAUSTION_WINDOW

• |NMD| still elevated, but TDR < 1 and REI improving.
• Reduce panic-taking; prefer improve/passive hybrid.

State D — TOXIC_CHASE

• |NMD| high and re-accelerating (PA up again), SBI spike, REI weak.
• Defensive: smaller clips, wider pauses, tail-protect controls.

State E — SAFE_FALLBACK

• Triggered by model uncertainty/drift/latency anomalies.
• Deterministic conservative schedule with hard slippage budget.

Use hysteresis to prevent flapping (entry threshold > exit threshold).

Branch-cost model

For each child-order decision (a \in {\text{join}, \text{improve}, \text{take}, \text{pause}}):

[ \mathbb{E}[C|a,s] = \underbrace{\mathbb{E}[\text{fill slippage}|a,s]}{\text{direct}} + \underbrace{\lambda_1,\mathbb{E}[\text{no-fill regret}|a,s]}{\text{completion}} + \underbrace{\lambda_2,\text{TailRisk}{q95}(a,s)}{\text{survival}} ]

where state (s) is from the regime machine above.

Key practical point: in EXHAUSTION_WINDOW, tail risk of immediate crossing often drops slower than expected benefit. This is where many naive urgency policies overtrade.

Control policy (practical)

1. PRESSURE_BUILD

   • Increase participation cap gradually.
   • Allow controlled taking.
   • Keep hard per-minute tail budget.

2. EXHAUSTION_WINDOW

   • Freeze further aggression escalation for a short TTL (e.g., 5–20s symbol-dependent).
   • Shift to improve/join mix.
   • Re-evaluate every short epoch; if PA re-accelerates, exit immediately.

3. TOXIC_CHASE

   • Clip size down.
   • Enforce cooldown between aggressive clips.
   • If budget burn-rate breaches threshold, escalate to SAFE_FALLBACK.

KPI pack

Track by symbol, session bucket, and regime:

• IS_bps (mean, p90, p95)
• Budget Burn Rate (realized / allowed slippage budget)
• Exhaustion Capture Rate (ECR): fraction of detected exhaustion windows where policy reduced next-interval markout vs baseline
• False Exhaustion Rate (FER): windows classified as exhaustion but followed by immediate toxicity re-acceleration
• Completion Reliability: residual notional at deadline

Promotion criteria (example):

• p95 IS improvement >= 8% vs baseline
• FER below fixed cap
• No degradation in completion reliability beyond tolerance

Data/engineering checklist

• Event-time synchronized L1/L2 + trades + child-order actions
• Deterministic reconstruction for backtest/replay
• Timestamp integrity checks (clock skew + feed latency flags)
• Feature freshness guardrails (stale features -> SAFE_FALLBACK)
• Per-state logging for post-trade attribution

Failure modes

1. Latency mirage: stale L2 creates fake exhaustion signal.
2. Regime leakage: thresholds tuned on calm periods only.
3. No hysteresis: rapid state flapping increases churn.
4. Objective drift: optimizing mean bps while p95 worsens.
5. Symbol overgeneralization: one threshold set across very different tick/queue regimes.

2-week rollout plan

Week 1

• Shadow-mode state labeling + KPI capture.
• No execution behavior change.
• Validate FER/ECR stability by session bucket.

Week 2

• Canary on small notional subset.
• Enable only EXHAUSTION_WINDOW aggression-freeze logic.
• Daily guardrail review (p95, budget burn, residual).

Rollback if:

• p95 cost worsens above threshold,
• completion drops below SLA,
• state flapping exceeds cap.

Bottom line

Microprice pressure is not binary. The last phase of pressure often behaves differently from the first. A state-aware exhaustion controller can reduce “panic-cross at local toxicity peak” behavior while keeping deadline risk bounded.

In short: model pressure lifecycle, not just pressure level.