Options-Expiry Gamma-Wall Unwind Slippage Playbook

Date: 2026-03-08 Category: research

Why this matters

Execution cost around options expiry is often mis-modeled as “just higher volatility.” In reality, dealer hedging flow can create strike-pinning (mean reversion toward large OI strikes) and then rapidly flip into unwind flow (directional acceleration after pin breaks or post-expiry de-hedging).

A static impact model calibrated on normal days will usually:

underprice urgency when a pin breaks,
overtrade into temporary pin stability (paying spread without real progress),
and miss tail slippage when gamma support vanishes.

This playbook treats expiry windows as a separate microstructure regime with explicit state control.

Core intuition

Near expiry, large open interest (OI) strikes act like local potential wells:

Short-gamma pressure low + balanced flow → price sticks near strike (“pin”).
Flow imbalance + local liquidity thinning → pin destabilizes.
Pin break + hedge rebalancing → realized impact becomes more persistent, with higher sweep risk.

So the execution question is not only “how much volatility?” but:

Are we in pinning, fragile pin, or unwind?
What is the probability of a pin-break transition during our execution horizon?
How does cost distribution (q50/q95) shift by state?

Data contract (minimum)

1) Spot + order book features

Mid, spread, top-N depth, imbalance, refill half-life
Cancel intensity, reject/retry frequency
Realized vol at short horizons (e.g., 10s/60s/5m)

2) Options-derived features

Strike-level OI concentration around spot (±2~3% moneyness)
Approx dealer gamma proxy by strike bucket
Time-to-expiry (minutes/hours)
Distance-to-nearest-high-OI strike (in ticks / bps)

3) Flow/transition features

Spot velocity away from candidate pin strike
Persistence of signed trade flow
Futures basis dislocation (if relevant for index names)
Auction proximity (open/close) flags

Regime labels (practical)

Define three operational states:

PINNED
- spot oscillates near high-OI strike
- low directional drift, higher mean reversion
- passive fill odds look good, but directional progress is slow
FRAGILE_PIN
- drift/flow imbalance increases
- spread/depth instability rises
- transition risk to break/unwind increases
UNWIND
- post-break directional movement
- impact persistence and sweep risk rise
- “wait for reversion” assumptions fail more often

Use a lightweight classifier (logit/GBM) for state probability vector:

[ P_t = (p_{pinned}, p_{fragile}, p_{unwind}) ]

with daily recalibration and strict monitoring of probability calibration.

Slippage model design

Model expected and tail cost jointly:

[ \hat C_{q}(x_t) = \sum_{s \in S} p_s(x_t) \cdot \hat C_{q}^{(s)}(x_t), \quad q \in {0.5, 0.95} ]

where each state-specific model includes:

spread + fee baseline,
transient impact term,
persistence term (larger in UNWIND),
no-fill / delay branch cost (critical in PINNED),
retry/reject tax (critical in FRAGILE_PIN/UNWIND).

Recommended branch decomposition

For each child decision (join/improve/take):

[ E[C] = P(fill)\cdot C_{fill} + (1-P(fill))\cdot C_{miss+catchup} ]

Do not train fills-only. Include miss/cancel branches to avoid survivorship bias.

Control policy (state-aware)

Use a simple 4-level action policy:

GREEN: normal passive bias
WATCH: moderate urgency bump, tighter dwell limits
STRESS: reduce passive dwell, increase controlled crossing
SAFE: cap participation, widen throttles, protect q95 budget

Suggested mapping

If p_unwind < 0.2 and q95 budget burn low → GREEN/WATCH
If p_fragile + p_unwind rises and depth half-life worsens → WATCH/STRESS
If pin-break probability spike + reject-loop metrics worsen → SAFE

Add hysteresis to prevent flip-flop.

Calibration and validation loop

Offline (weekly)

Refit state classifier + state-conditional cost models
Reliability plots for state probabilities
Quantile coverage tests (q50/q95)
Slice-by-slice stress checks (expiry day vs normal)

Online (intraday)

Drift in feature distribution (PSI/KS)
Calibration drift of state probabilities
Budget-burn monitor:

[ Burn_t = \frac{RealizedCost_t}{PlannedBudget_t} ]

Escalate control level when burn-rate breaches thresholds.

Execution heuristics that usually help

Near strong pin, avoid fake progress chasing
- passive fills may be high but directional completion weak.
When fragile-pin signals rise, shorten dwell time
- queue age becomes less valuable; reset tax can explode.
After confirmed break, avoid binary panic sweeps
- move to staged aggression with hard tail budget gates.
Separate expiry-day parameter set
- do not rely on non-expiry pooled calibration.

Rollout plan

Shadow mode (2–4 weeks)
- produce state probs + recommended control action without execution changes.
Canary (5–10% flow)
- only allow GREEN/WATCH decisions initially.
Expanded canary
- enable STRESS/SAFE with strict rollback criteria.
Full deploy with champion/challenger
- keep baseline controller as automatic fallback.

Rollback triggers

q95 slippage breach for N consecutive windows
no-fill regret exceeds threshold
classifier calibration collapse (reliability failure)

Minimal KPI dashboard

State probabilities over time (p_pinned, p_fragile, p_unwind)
Pin-break event count + detection latency
q50/q95 slippage vs baseline
Fill ratio and no-fill regret by state
Retry/reject tax by state
Budget burn-rate and control-state occupancy

Common failure modes

Treating expiry effects as plain volatility scaling
Overfitting to one expiry cycle
Ignoring options-derived features due to data friction
Using instantaneous state labels without transition hysteresis
Evaluating only average bps (tail blind)

Bottom line

On expiry windows, execution quality depends on state transitions (pin → fragile → unwind), not just spread/vol snapshots.

A practical edge comes from:

explicit regime probabilities,
state-conditional tail-cost modeling,
and budget-linked control escalation.

That is how you prevent “looked fine at q50” from becoming expensive at q95 when gamma support disappears.