Venue-Specific Impact-Decay Calibration for Live Slippage Control

Date: 2026-03-03
Category: research
Focus: Practical modeling framework to separate temporary vs. persistent impact by venue and convert it into routing/participation controls.

Why this matters

Most desks estimate one global slippage curve, then wonder why costs explode when routing mix changes. Impact decay is not universal:

• queue refill speed differs by venue
• maker/taker incentives alter adverse-selection pressure
• cancel intensity and hidden-liquidity behavior change the relaxation path

If you fit one decay kernel for all venues, you overtrade “slow-decay” venues and underuse “fast-decay” venues.

Modeling target

For parent order slices indexed by (i), model signed future price move after each child fill:

[ \Delta p_{i,\tau}=\alpha_{v,s,t}+\beta_{v,s,t},q_i^{\delta},G_v(\tau)+\Gamma_{s,t},M_i+\epsilon_{i,\tau} ]

• (v): venue, (s): symbol, (t): regime bucket (vol/liquidity/time)
• (q_i): signed child size (or participation-normalized size)
• (G_v(\tau)): venue-specific decay kernel (temporary impact relaxation)
• (M_i): controls (spread, imbalance, queue position proxy, short-horizon volatility)

Use a two-part view:

1. Transient component (decays): execution footprint you may recover by waiting
2. Persistent component (does not decay quickly): likely information/toxicity cost

Calibration pipeline (desk-operational)

1. Event table: child fill timestamp, side, size, venue, fee/rebate, queue metrics, top-of-book state.
2. Horizon grid: (\tau\in{1s,5s,15s,30s,60s,180s}) markouts (mid-based + executable-based).
3. Regime stratification: volatility terciles × spread state × session bucket.
4. Kernel family per venue:
   • power-law: (G(\tau)=(1+\tau/\theta)^{-\kappa})
   • stretched exp: (G(\tau)=\exp(-(\tau/\theta)^\nu))
5. Hierarchical shrinkage: venue-symbol cells borrow strength from venue-level priors to avoid noisy small-sample fits.
6. Walk-forward refit: daily incremental + weekly full refit; freeze parameters intra-session except alert-triggered fallback.

Key metrics to store

• Half-life of temporary impact (t_{1/2,v})
• Persistent fraction (\pi_v = \text{impact}(180s)/\text{impact}(1s))
• Recovery uncertainty band (P10/P50/P90)
• Net impact after fees/rebates (must be routing objective, not raw markout)

Converting model to execution decisions

1) Venue routing score

[ \text{Score}_v = -\widehat{\text{NetCost}}_v + w_r,\widehat{\text{Recovery}}_v - w_u,\widehat{\text{Uncertainty}}_v ] Route marginal flow to highest score under risk caps.

2) Participation throttle

If realized recovery < model P10 for N consecutive windows, reduce POV cap (e.g., 12% → 8%) and widen inter-slice gap.

3) Cooldown logic

For slow-decay venues (high (t_{1/2}), high (\pi)), add mandatory cooldown between aggressive clips, unless urgency state is critical.

Monitoring & drift alarms

Trigger alerts when any holds for >15 min:

• realized 30s markout error z-score > 3
• half-life estimate shift > 40% vs trailing 20-day baseline
• persistent fraction jump with concurrent cancel-rate spike (possible toxicity regime change)

Fallback policy:

• switch to conservative kernel prior
• clamp aggression and reweight toward historically fast-recovery venues
• log incident tag for post-trade TCA review

Failure modes (seen in production)

• Selection bias: only modeling filled children, ignoring unfilled opportunities.
• Clock skew: venue timestamp alignment errors create fake decay patterns.
• Fee blindness: positive raw markout but negative net economics after fees.
• Overfit horizons: too many (\tau) points without shrinkage yields unstable controls.

Minimum implementation checklist

• Drop-copy + market data join validated (ms-level clock QA)
• Executable markout path (not only mid)
• Venue-specific kernel with hierarchical pooling
• Net-cost objective includes fee/rebate + spread capture
• Online drift monitor + automatic safe fallback
• Weekly champion/challenger review with rollback criteria

References

• Almgren, R., & Chriss, N. (2000). Optimal Execution of Portfolio Transactions.
• Bouchaud, J.-P. et al. (propagator/transient impact literature).
• Taranto, D. et al. (2016). Linear models for the impact of order flow on prices I. Propagators (arXiv:1602.02735).

Bottom line: model impact decay by venue, not globally. The edge is not just lower average slippage; it is faster detection of regime breaks and safer automatic throttling before costs compound.